Schrödinger Equations

Solutions to the Nonlinear Schrdinger Equation
Carrying Momentum along a Curve

FETHI MAHMOUDI
La Facult des Sciences de Tunis
ANDREA MALCHIODI
SISSA
AND
MARCELO MONTENEGRO
Universidade Estadual de Campinas
Abstract
We prove existence of a special class of solutions to the (elliptic) nonlinear
Schrdinger equation c
2
^[ V(.)[ = [[[
;1
[ on a manifold or in Eu-
clidean space. Here V represents the potential, ] an exponent greater than 1,
and c a small parameter corresponding to the Planck constant. As c tends to 0
(in the semiclassical limit) we exhibit complex-valued solutions that concentrate
along closed curves and whose phases are highly oscillatory. Physically these
solutions carry quantum-mechanical momentum along the limit curves. 2008
Wiley Periodicals, Inc.
1 Introduction
In this paper we are concerned with concentration phenomena for solutions of
the singularly perturbed elliptic equation
(NLS
t
) c
2
^
v
[ V(.)[ = [[[
;1
[ on M.
where M is an n-dimensional compact manifold (or the at Euclidean space R
n
),
V a smooth, positive function on M satisfying the properties
(1.1) 0 < V
1
_ V _ V
2
. |V |
C
3 _ V
3
(for some xed constants V
1
. V
2
, and V
3
), [ a complex-valued function, c > 0
a small parameter, and ] an exponent greater than 1. Here ^
v
stands for the
Laplace-Beltrami operator on (M. g).
Equation (NLS
t
) arises from the study of the focusing nonlinear Schrdinger
equation
(1.2) i
d

[
dt
=
2
^

[ V(.)

[ [

[[
;1
[ on M 0. o).
Communications on Pure and Applied Mathematics, Vol. LXII, 11551264 (2009)
2008 Wiley Periodicals, Inc.
1156 F. MAHMOUDI, A. MALCHIODI, AND M. MONTENEGRO
where

[ =

[(.. t ) is the wave function, V(.) a potential, and the Planck con-
stant. A special class of solutions to (1.2) consists of functions whose dependence
on the variables . and t are of the form

[(.. t ) = e
iot
[(.). Such solutions
are called standing waves and up to substituting V(.) with V(.) o, they give
rise to solutions of (NLS
t
) for c = .
An interesting case is the semiclassical limit c 0, where one should expect
to recover the Newton law of classical mechanics. In particular, near stationary
points of the potential, one is led to search highly concentrated solutions, which
could mimic point particles at rest.
In recent years, much attention has been devoted to the study of the above prob-
lem: one of the rst results in this direction was due to Floer and Weinstein [21],
where the case of M = R and ] = 3 was considered, and where existence of
solutions highly concentrated near critical points of V has been proved. This result
has since been extended by Oh [48] to the case of R
n
for arbitrary n, provided
1 < ] <
nC2
n2
. The prole of these solutions is given by the ground state U
x
0
(namely, the solution with minimal energy, which is real-valued, positive, and ra-
dial) of the limit equation
(1.3) û V(.
0
)u = u
;
in R
n
.
where .
0
is the concentration point. The solutions u
t
obtained in the aforemen-
tioned papers behave qualitatively like
u
t
(.) . U
x
0
. .
0
c
as c tends to 0, and since U

x
0
decays exponentially to 0 at innity, u
t
vanishes
rapidly away from .
0
.
The above existence results have been generalized in several directions, includ-
ing the construction of solutions with multiple peaks, the case of degenerate poten-
tials, potentials tending to 0 at innity, and more general nonlinearities. We refer
the interested reader to the (incomplete) list of works [1, 2, 3, 6, 7, 8, 13, 17, 23, 28]
and to the bibliographies therein.
We also mention the mathematical similarities between (NLS
t
) and problem
(1
t
)
c
2
û u = u
;
in C.
Ju
J
= 0 on dC.
u > 0 in C.
where C is a smooth bounded domain of R
1
, ] > 1, and v denotes the exterior
unit normal vector to dC. Problem (1
t
) arises in the study of some biological mod-
els; see, e.g., [44] and references therein, and as (NLS
t
), it exhibits concentration of
solutions at some points of

C. Since the last equation is homogeneous, the location
of the concentration points is determined by the geometry of the domain. About
this topic, we refer the reader to [14, 18, 24, 25, 26, 27, 31, 32, 33, 45, 46, 47, 52].
SOLUTIONS TO NLS CARRYING MOMENTUM ALONG A CURVE 1157
More recently new types of solutions to (NLS
t
) have been found, since when c
tends to 0 these solutions do not concentrate at points but instead at sets of higher
dimension. Before stating our main result, it is convenient to recall the progress
on this topic and to illustrate the new phenomena involved. Some rst results
in this direction were given in [9, 11] in the case of radial symmetry, and later
improved in [4] (see also [5] for the problemin bounded domains), where necessary
and sufcient conditions for the location of the concentration set have been given.
Unlike the previous case, the limit set is not stationary for the potential V : indeed,
from heuristic considerations, the energy of a solution concentrated near a sphere
of radius r depends both on V and on its volume, proportional to cr
n1
.
Based on the above energy considerations, in [4] a conjecture is stated concern-
ing concentration on k-dimensional manifolds for k = 1. . . . . n 1; it is indeed
expected that, under suitable nondegeneracy assumptions, the limit set should sat-
isfy the equation
(1.4) 0
k
V
1
V = V H with 0
k
=
] 1
] 1
1
2
(n k).
where V
1
stands for the normal gradient and H for the curvature vector, and the
prole of the solutions at a point .
0
in the limit set should be asymptotic, in the
normal directions, to the ground state of
(1.5) û V(.
0
)u = u
;
in R
nk
.
Since the Pohozaev identity implies ] <
nkC2
nk2
for the existence of nontrivial
solutions, the latter condition is expected to be a natural one for dealing with this
phenomenon.
Actually, concerning (1
t
) another conjecture has been stated, asserting exis-
tence of solutions concentrating at sets of positive dimension. About the latter
problem, starting from the paper [40], there has been some progress in the general
setting (without symmetry assumptions), and after the works [34, 39, 41], existence
is now known for arbitrary dimension and codimension. About problem (NLS
t
),
the conjecture in [4] has been veried in [19] for n = 2 and k = 1. Some other
(and related) results, under some reduced symmetry assumptions have been given
in [10, 15, 43, 49].
It is worth pointing out a major difference between the symmetric and the non-
symmetric situation. In fact, since the ground states of (1.3) or (1.5) are of moun-
tain-pass type (namely critical points of some Euler functional with Morse index
equal to 1), equation (NLS
t
) becomes highly resonant. To explain this phenome-
non, we consider a real-valued function [ in R
2
with a radial potential. We can
begin by nding approximate (radial) solutions of the form u
N i,t
(r) . U
N i
(
iN i
t
),
where U
N i
is the solution of (1.3) for n = 1 corresponding to V( r). Then, with a
good choice of r, we can try to linearize the equation and nd true solutions via a
local inversion. The linearized equation, taking [ real for simplicity, becomes
c
2
^[ V(r)[ ]u
N i,t
(r)
;1
[ in R
2
.
Using polar coordinates (r. 0) and a Fourier decomposition of [ with respect to 0,
[(r. 0) =

}
e
i}
[
}
(r), we see that the following operator acts on each compo-
nent [
}
:
(1.6) c
2
[
00
}
c
2
1
r
[
0
}
V(r)[
}
]u
N i,t
(r)
;1
[
}

1
1;"
v
j
1
r
2
c
2
;
2
[
}
on 0. o).
where 1
1,t
(apart from the term c
2 1
i
[
0
}
, which is not relevant to the next discus-
sion) represents the linearized equation of (NLS
t
) in one dimension near a soliton.
Since one expects to deal with functions that are highly concentrated near r = r,
the last term in the above formula naively increases the eigenvalues by a quantity
of order
1
N i
2
c
2
;
2
compared to those of 1
1,t
.
The operator 1
1,t
possesses a negative eigenvalue n
t
lying between two nega-
tive constants independent of c (since U
N i
is of mountain-pass type, as explained
before) and a (nearly) zero eigenvalue o
t
, by the translation invariance of (1.3) in
R
1
. As a consequence, the operator in (1.6) possesses two sequences of eigenval-
ues qualitatively of the form n
},t
. n
t
c
2
;
2
and o
},t
. o
t
c
2
;
2
. This might
generate two kinds of resonances: for small values of ; , when o
},t
. 0, and for ;
of order
1
t
, when n
},t
could be close to 0.
A comment is in order on resonant modes, which can be roughly studied with a
separation of variables as before. The ones relative to o
},t
(for ; small) oscillate
slowly along the limit set, while the ones relative to the resonant n
},t
s oscillate
quickly with the number of oscillations proportional to k .
1
t
.
The invertibility of the linearized operator will then be equivalent to having all
the o
},t
s and all the n
},t
s different from 0. A control on the resonant o
},t
s can be
obtained (via some careful expansions) from a suitable nondegeneracy condition
involving the limit set and the potential V . On the other hand, the possible van-
ishing of some n
},t
is peculiar to this concentration behavior and more intrinsic, so
invertibility can only be achieved by choosing suitable values of c.
These formal considerations can also apply to the case of concentration near a
general manifold (without symmetries) in higher dimension or codimension, and
related phenomena appear in some geometric problems as well that deal with the
construction of surfaces with constant mean curvature; see [38, 42]. When C is a
radially symmetric domain and the potential V is radially symmetric, the problem
is simpler, since working in spaces of invariant functions avoids most of the above
resonances.
In this paper we construct a newtype of solution, which concentrates along some
curve ,, and which physically carries momentum along the limit set. Differently
from those discussed before, these solutions are complex valued and their prole
near any point .
0
in the image of , is asymptotic to a solution to (1.3), which
decays exponentially to 0 away from the .
n
-axis of R
n
and is periodic in .
n
. More
precisely, we consider proles of the form
(1.7) (.
0
. .
n
) = e
i
y
( x
n
U(.
0
). .
0
= (.
1
. . . . . .
n1
).
where

} is some constant and

U(.
0
) a real function. With this choice of , the
function

U satises
(1.8) ^
U (

}
2
V(.
0
))
U = [

U[
;1
U in R
n1
and decays to 0 at innity. Solutions to (1.8) can be found by considering the (real)
function U satisfying Û U = U
;
in R
n1
(decaying to 0 at innity), and
by using the scaling
(1.9)
U(.
0
) =

hU(
k.
0
).
h = (

}
2
V(.
0
))
1
p1
.

k = (

}
2
V(.
0
))
1
2
.
In the above formulas

} can be taken arbitrarily, and

h and

k have to be chosen ac-
cordingly, depending on V(.
0
). The constant

} represents the speed of the phase
oscillation and is physically related to the velocity of the quantum-mechanical par-
ticle described by the wave function.
Usually standing waves have zero angular momentum, and with the exception
of the present paper we are aware of only one result in this direction, given in [16]
(see also the comments there) where the case of an axially symmetric potential
is considered. Our goal here is to treat this phenomenon in a generic situation,
without any symmetry restriction. Some of the difculties of such an extension
were naively summarized in the above discussion, but some new ones arise due
to the fact that the standing waves are complex valued and their phase is highly
oscillatory; more comments on these issues are given later.
Before stating our main result, we discuss how to determine the limit set. If we
look for a solution [ to (NLS
t
) with the above prole, then it should qualitatively
behave as
(1.10) [( s. ) . e
i
f.N s/
"
h( s)U
k( s)
c
.
where s is the arc length parameter of ,, and a system of geodesic coordinates
normal to ,. To obtain more exibility, we choose the phase oscillation to depend
on the point ,( s), while h( s) and k( s) should satisfy
(1.11) h( s) =

(}
0
( s))
2
V(,( s))
1
p1
. k( s) =

(}
0
( s))
2
V(,( s))
1
2
.
which is the counterpart of (1.9) for a variable potential.
The function }( s) can be (heuristically) determined by using an expansion of
(NLS
t
) at order c; a computation performed in Subsection 2.3 (see in particular
formula (2.8)) shows that
(1.12) }
0
( s) . Ah
c
( s) with o =
(n 1)(] 1)
2
2.
where A is an arbitrary constant. At this point, only the curve , should be deter-
mined.
First of all, we notice that the phase should be a periodic function in the length
of the curve, and therefore by (1.12) it is natural to work in the class of loops
(1.13) I :=
, : R M periodic : A
,
h( s)
c
J s = const
.
where s stands for the arc length parameter on ,. Problem (NLS
t
) has a variational
structure, with the Euler-Lagrange functional given by
1
t
([) =
1
2
(c
2
[V
v
[[
2
V(.)[[[
2
)
1
] 1
[[[
;C1
.
For a function of the form (1.10), by a scaling argument (see (2.11)) we have
(1.14) 1
t
([) . c
n1
,
h( s)
0
J s with 0 = ] 1
1
2
(] 1)(n 1):
therefore a limit curve , should be a critical point of the functional

,
h( s)
0
J s in
the class I. With a direct computation (see Subsection 2.4), we can check that the
extremality condition is
(1.15) V
1
V = H
] 1
0
h
;1
2A
2
h
2c
where, as before, V
1
V represents the normal gradient of V and H the curvature
vector of ,. Similarly, via some long but straightforward calculation, we can nd
a natural nondegeneracy condition for stationary points that is expressed by the
invertibility of the operator in (2.22) acting on the normal sections to , (we refer
the reader to Section 2 for the notation used in the formula). We notice that, since
formula (1.12) determines only the derivative of the phase, to obtain periodicity we
need to introduce some nonlocal terms; see (2.14). After these preliminaries, we
are in position to state our main result.
THEOREM 1.1 Let M be a compact n-dimensional manifold, let V : M R be
a smooth, positive function, and let 1 < ] <
nC1
n3
. Let 1 > 0; then there exists
a positive constant A
0
, depending on V , ], and 1, for which the following holds:
If 0 _ A < A
0
, if , has length less than or equal to 1 and satises (1.15), and
the operator in (2.22) is invertible on the normal sections of ,, there is a sequence
c
k
0 such that problem (NLS
t
k
) possesses solutions [
t
k
having the asymptotics
in (1.10), with } satisfying (1.12).
Remark 1.2.
(i) The statement of Theorem 1.1 remains unchanged if we replace M by R
n
(or with an open manifold asymptotically Euclidean at innity), and we assume
V to be bounded between two positive constants and for which |V
/
V | _ C
/
,
l = 1. 2. 3, for some positive constants C
/
.
(ii) The restriction on the exponent ] is natural, since (1.8) admits solitons if
and only if ] is subcritical with respect to dimension n 1.
(iii) The smallness requirement on Ais technical, and we believe this condition
can be relaxed. Anyway, for
nC2
n2
_ ] <
nC1
n3
, A should have an upper bound
depending on V to have solvability for both (1.9) and (1.12). About this condition,
see Remark 2.2 and Remark 4.7.
(iv) Apart from the assumption on A, Theorem 1.1 improves the result in [16].
In fact, in addition to removing the symmetry condition (which is the main issue),
the characterization of the limit set is explicit, the assumptions on V are purely
local, and the upper bound on ] is sharp.
(v) The existence of solutions to (NLS
t
) only for a suitable sequence c
k
0
is related to the resonance phenomenon described above. The result can be ex-
tended to a sequence of intervals in the parameter c approaching 0 but, at least with
our proof, we do not expect to nd existence for all epsilons.
Taking A = 0 (hence }
0
0), from (1.11) it follows that V = h
;1
and that
(1.15) is equivalent to (1.4), so as a consequence of our theorem, we can prove the
conjecture in [4] for k = 1, extending the result in [19].
COROLLARY 1.3 Let M be a compact Riemannian n-dimensional manifold with
metric g, let V : M R be a function satisfying (1.1), and let 1 < ] <
nC1
n3
.
Let , be a simple closed curve that is a nondegenerate geodesic with respect to the
weighted metric
V
pC1
p1
n1
2
g.
Then there exists c
k
0 such that problem (NLS
t
k
) has real-valued solutions
[
t
k
concentrating near , as ; o and having the asymptotics
[
t
k
( s. ) . V(,( s))
1
p1
U
V(,( s))
1
2
c
k
.
where s stands for the arc length parameter of ,, and for a system geodesic
coordinates normal to ,.
Corollary 1.3 also gives some criterion for the applicability of Theorem 1.1; in
fact, starting froma nondegenerate geodesic in the weighted metric, via the implicit
function theorem for A sufciently small, we obtain a curve for which (1.15) and
the invertibility of (2.22) hold. In particular, when V is constant, we can start with
nondegenerate close geodesics on M in the ordinary sense.
To prove Theorem 1.1 we proceed as follows: We collect some preliminary
material in Section 2, where we recall some geometric facts, and we study the
functional in (1.14) constrained to the class of curves I, determining the Euler-
Lagrange equation together with the nondegeneracy condition.
In Section 3 we derive some expansions (in powers of c) of equation (NLS
t
) for
[ of the form (1.10). To do this, it is convenient to scale problem (NLS
t
) in the
following way:
(1.16) ^
v
"
[ V(c.)[ = [[[
;1
[ in M
t
.
where M
t
denotes the manifold M endowed with the scaled metric g
t
= (1c
2
)g
(with an abuse of notation we might often write M
t
=
1
t
M, and if . M
t
, we
write c. to indicate the corresponding point on M).
We are now looking for a solution concentrated near the dilated curve ,
t
:=
1
t
,.
We let s be the arc length parameter of ,
t
so that s = cs, and we let (1
}
)
}D2,...,n
denote an orthonormal frame in N, (the normal bundle of ,) transported in parallel
to the normal connection; see Section 2. We also let (.
}
)
}
be a corresponding set
of normal coordinates. Since we want to allow some exibility both in the choice
of the phase and of the curve of concentration, we dene

}
0
( s) = }( s) c}
1
( s),
and we set z
}
= .
}

}
( s), where (
}
)
}D2,...,n
are the components (with respect
to the coordinates .) of a section in N,. Then, with a formal expansion of [ in
powers of c up to second order, in the coordinates (s. z) near ,
t
, we set
[
2,t
(s. z) = e
i
z
f
0
."s/
"
{h(cs)U(k(cs)z) cn
i
in
i
| c
2
i
i
i
|].
s 0.
1
t
|, z R
n1
, and 1 = 1(,), the length of ,, for some corrections n
i
,
n
i
,
i
, and
i
(which have to be determined) to the above approximate solutions.
In Subsection 3.2 we show that these terms satisfy equations of the form L
i
n
i
= F
i
, L
i
n
i
= F
i
, L
i
i
=

F
i
, and L
i
i
=

F
i
where
(1.17)
L
i
= ^
:
V( s) ]h( s)
;1
U(k( s)z)
;1
L
i
= ^
:
V( s) h( s)
;1
U(k( s)z)
;1
in R
n1
.
and where F
i
, F
i
,

F
i
, and

F
i
are given data that depend on V , ,, s, A, , and
}
1
. The operators L
i
and L
i
are Fredholm (and symmetric) from H
2
(R
n1
) into
1
2
(R
n1
), and the above equations for the corrections can be solved provided the
right-hand sides are orthogonal to the kernels. It is well-known (see, e.g., [30]) that
L
i
has a single negative eigenvalue, a kernel with multiplicity n 1 spanned by
the functions d
/
U(k( s)z), l = 2. . . . . n (the generators of the normal translations),
while all the remaining eigenvalues are positive. The operator L
i
instead has one
zero eigenvalue with eigenfunction U(k( s)z) (the generator of complex rotations)
and all the remaining eigenvalues positive. As explained above, condition (1.15)
and the nondegeneracy of the operator J help us to determine n
i
, n
i
,
i
, and
i
(and }
1
), respectively, namely to solve (1.16) at order c rst, and then at order c
2
.
As discussed before, some fast-oscillating functions (along ,) contribute to gen-
erate some resonance, but these aspects are treated in later sections. Even at this
stage, however, there are new difculties compared to the results in [19, 34, 39, 40].
In our case the solutions are complex valued, and this causes an extra degener-
acy due to their invariance under multiplication by a phase factor. As a conse-
quence, we have a further (innite-dimensional) approximate kernel, correspond-
ing roughly to a factor of [
t
in the form e
i(
1
(N x)
for }
1
arbitrary. The correction
in the phase can also be determined by a formal expansion in c and, as for }
0
,
we still obtain nonlocal terms. Also, when expanding formally the solutions in
c, the highly oscillatory behavior of solutions generates an increasing number of
derivatives in s.
In Section 4 then we set up the strategy to obtain true solutions from the approx-
imate ones. First of all, since U (and its derivatives) decay fast at innity like
(1.18) U(r) . e
i
r
n2
2
as r o.
it is possible to localize the problem in a neighborhood of the scaled curve ,
t
; this
step is inspired by [19] and worked out in Subsection 4.2. We then try to nd a true
solution of the form
e
i
z
f ."s/
"
{h(cs)U(k(cs)z) cn
i
in
i
| c
2
i
i
i
| n].
with n suitably small and

} close to

}
0
, via some local inversion arguments. From
a linearization of the equation near [
2,t
, the operator 1
t
acting on n in the coor-
dinates (s. z) is then the following:
1
t
n := d
2
xx
n ^
:
n V(c.) [[
2,t
[
;1
n
(] 1)[[
2,t
[
;3
[
2,t
m([
2,t

n).
(1.19)
Here mdenotes the real part. Decomposing rst n into its real and imaginary parts,
and then in Fourier modes with respect to the variable cs, we can write
n = n
i
i n
i
=
}
sin(;cs) n
i,}
(z) i
}
sin(;cs) n
i,}
(z)
(forgetting for simplicity about the cosine functions). If we take (as a model prob-
lem) V 1, then the operators (in the z-variables) acting on the real and imaginary
components are, respectively, L
i
c
2
;
2
and L
i
c
2
;
2
.
As for (1.6), the kernels of L
i
and L
i
produce a sequence of eigenvalues for
1
t
that behave qualitatively like c
2
;
2
, and for small values of ; these become
resonant. With an accurate expansion of these eigenvalues, we nd that the non-
degeneracy assumption on (2.22) prevents each of them from vanishing. However,
the fact that L
i
possesses a negative eigenvalue as well generates an extra sequence
of eigenvalues of 1
t
, qualitatively of the form 1 c
2
;
2
, ; N. As explained
before, the only hope to get invertibility is to choose the values of c appropriately.
For the Neumann problem (1
t
) the resonance phenomenon was taken care of us-
ing a theorem by T. Kato [29, p. 445], which allows us to differentiate eigenvalues
with respect to c. In the aforementioned papers it was shown that when varying the
parameter c the spectral gaps near 0 shrink only slightly, and invertibility can be
obtained for a large family of epsilons.
However, when the concentration set is one dimensional, the spectral gaps of
the resonant eigenvalues (with fast-oscillating eigenfunctions) are relatively large,
of order c, and the prole of the corresponding eigenfunctions can be analyzed by
means of a scalar function on 0. 1| (see below and in Subsection 6.2). This might
allow us to bypass Katos theorem and use a more direct approach, employed in
[42] to exhibit constant mean curvature surfaces of cylindrical type embedded in
manifolds, and in [19] for studying solutions of (NLS
t
) in R
2
. We can partially
take advantage of these techniques (see the comments in Section 4), but some new
difculties arise due to the phase oscillations in (1.10).
From the above discussion, we expect to nd three possible resonances: two
of them for small values of the index ; (with eigenvectors roughly of the form
e
i(((tx))t
d
/
U(k( s)z) sin(c;s), l = 2. . . . . n, and i e
i(((tx))t
U(k( s)z) sin(c;s),
respectively), and a third one for ; of order
1
t
, precisely when c
2
;
2
coincides
with the rst eigenvalue of L
i
.
To understand this behavior, we rst study the spectrum of a model operator
similar to (1.19), where we assume V

V > 0 and [
2,t
to coincide with the
function in (1.7). For this case we characterize completely the spectrum of the op-
erator and the properties of the eigenfunctions; see Subsection 4.3 and in particular
Proposition 4.5. The condition on the smallness of A appears precisely here (and
only here) and is used to show that the resonant eigenvalues are only of the forms
described above. Removing the smallness assumption might indeed lead to further
resonance phenomena; see Remark 4.7 for related comments.
We next consider the case of nonconstant potential V . Since this has a slow
dependence in s along ,
t
, one might guess that the approximate kernel of 1
t
(see
(1.19)) might be obtained from that for constant V , allowing a slow dependence
in s of the prole of these functions. With this criterion, given a small positive
parameter , we introduce a set 1
(see (4.50) and the previous formulas) consist-

ing of candidate approximate eigenfunctions on 1
t
, once multiplied by the phase
factor e
i(((tx))t
. More comments on the specic construction of this set can be
found in Subsection 4.3, especially before (4.50).
In Proposition 4.9 we show that this guess is indeed correct; in fact, we prove
that the operator 1
t
is invertible provided we restrict ourselves to the subset

H
t
of
functions that are orthogonal to e
i(((tx))t
1
. This property allows us to solve

the equation up to a Lagrange multiplier in 1
; see Proposition 4.14. For tech-

nical reasons, we prove invertibility of 1
t
in suitable weighted norms, which are
convenient to deal with functions decaying exponentially away from ,
t
.
Compared to the other papers that deal with this kind of resonance, the approx-
imate kernel here depends genuinely on the variable s (in [34, 39, 40, 41, 42]; the
problem is basically homogeneous along the limit set, while in [19] it can be made
such through a change of variables). To deal with this feature, which is the main
cause of difculty in Proposition 4.9, we localize the problem in the variable s
as well. Multiplying by a cutoff function in s, we show that orthogonality to 1

implies approximate orthogonality to the set

1
; see (4.53) and the previous for-

mulas, which is the counterpart of 1
for a potential frozen at some point in ,

t
;
once this is shown, we use the spectral analysis of Proposition 4.5.
Section 5 is devoted to choosing a family of approximate solutions to (1.16);
since we have many small eigenvalues appearing, it is natural to try looking for
functions that solve (1.16) as accurately as possible. Our nal goal is to annihilate
the Lagrange multiplier in Proposition 4.14, and to do this we choose approximate
solutions

2,t
(in the notation of Section 5), which depend on suitable parameters:
a normal section , a phase factor }
2
, and a real function . These parameters cor-
respond to different components of 1
and are related to the kernels of L

i
(c
2
;
2
)
and L
i
(c
2
;
2
); see the above comments. The function in particular is highly
oscillatory and takes care of the resonances due to the fast Fourier modes.
We next need to derive rigorous estimates on the error terms and to study in
particular their Lipschitz dependence on the data , }
2
, and . Proposition 5.2
collects the nal expression for ^
v
"
[ V(c.)[ [[[
;1
[ on the approximate
solutions

2,t
; the error terms

A are listed (and estimated) before in that section,
together with their Lipschitz dependence on the parameters.
Finally, after performing a Lyapunov-Schmidt reduction onto the set 1
(see
Proposition 6.1), we study the bifurcation equation. In doing this we crucially
use the formal computations in Section 3 and the error estimates in Section 5. In
particular, for and }
2
we nd as main terms, respectively, the operator J in
(2.22) and the one in the middle of (3.24), both appearing when we perform formal
expansions. These operators are both invertible by our assumptions, and therefore
we are able to determine and }
2
without difculty.
The operator acting on instead is more delicate, since it is qualitatively of the
form
(1.20) c
2
00
( s) z( s) on 0. 1|.
with periodic boundary conditions, where z is a negative function. This operator is
precisely the one related to the peculiar resonances described above. In particular,
it is resonant at frequencies of order
1
t
, and this requires us to choose a norm for
that is weighted in the Fourier modes; see (5.6) and Subsection 6.2. For operators
like that in (1.20) there is in general a sequence of epsilons for which a nontrivial
kernel exists. Using Katos theorem though, as in [34, 38, 39, 40, 41], we pro-
vide estimates on the derivatives of the eigenvalues with respect to c, showing that
for several values of this parameter the operator acting on is invertible. In this
operation also the value of the constant A (see (1.12)) has to be suitably modi-
ed (depending on c) in order to preserve the periodicity of our functions. Once
we have this, we apply the contraction mapping theorem to solve the bifurcation
equation as well.
The results in this paper were rst written in two different preprints [35, 36].
Some lengthy proofs, consisting of many explicit computations, have been omitted
here for reasons of brevity, but precise references to the preprints will be given.
Theorem 1.1 was announced in the note [37].
Notation and Conventions
v Dealing with coordinates, capital letters like . T. . . . will vary between 1
and n while indices like ;. l. . . . will run between 2 and n. The symbol i
will always stand for the imaginary unit.
v For summations, we use the standard convention of summing terms where
repeated indices appear.
v We will choose coordinates (.
1
. . . . . .
n
) near a curve ,, and we will pa-
rametrize , by arc length letting .
1
= s. Its dilation ,
t
:=
1
t
, will be
parametrized by s =
1
t
s. The length of , is denoted by 1.
v For simplicity, a constant C is allowed to vary from one formula to another,
and also within the same line.
v For a real positive variable r and integer m, O(r
n
) (respectively, o(r
n
))
will denote a complex-valued quantity for which [O(r
n
)r
n
[ remains
bounded (respectively, [o(r
n
)r
n
[ tends to 0) when r tends to 0. We might
also write o
t
(1) for a quantity that tends to 0 as c tends to 0.
2 Study of the Reduced Functional
In this section we consider the functional in the right-hand side of (1.14) dened
on the set I, representing the approximate energy 1
t
of a function concentrated
near , with the prole (1.10). We rst introduce a convenient set of coordinates
near an arbitrary (smooth) closed curve in M. Then, using these coordinates, we
write the Euler equation and the second variation formula at a stationary point.
2.1 Geometric Preliminaries
In this subsection we discuss some preliminary geometric facts, referring, for
example, to [20, 50]. Given an arbitrary simple closed curve , in M, we choose
coordinates .
1
. . . . . .
n
near ,, called Fermi coordinates, in the following way: We
let .
1
parametrize the curve , by arc length. At some point q in the image of ,, we
consider an orthonormal (n 1)-tuple (Y
2
. . . . . Y
n
), which forms a basis for N
q
,,
the normal bundle of , at q. We extend the Y
/
s as vector elds along , via parallel
transport along the curve with respect to the normal connection V
1
, namely, by
the condition V
1
P ,
Y
/
= 0 for l = 2. . . . . n.
Next we parametrize a point near , using the following coordinates ( s. .)
R R
n1
,
( s. .
2
. . . . . .
n
) exp
,(N x)
(.
2
Y
2
.
n
Y
n
).
where exp
q
is the exponential map in M through the point q. In this way, xing s,
each curve t t., for . R
n1
\ {0] and t close to 0, is mapped into a geodesic
in M passing through ,( s).
Let us now dene the vector elds 1
1
=
J
JN x
and 1
/
=
J
J+
l
for l = 2. . . . . n.
We notice that on , each 1
/
coincides with Y
/
, while 1
1
on , is nothing but ,.
By our choice of coordinates it follows that V
T
1 = 0 on , for any vector eld 1
that is a linear combination (with coefcients depending only on s) of the 1
}
s,
; = 2. . . . . n. In particular, for any l. ; = 2. . . . . n and for any R, we have
V
T
l
CT
j
(1
/
1
}
) = 0 on ,, which implies V
T
l
1
}
V
T
j
1
/
= 0 for every
l. ; = 2. . . . . n. Using the fact that 1
s are coordinate vectors for = 1. . . . . n

and in particular V
T
A
1
B
= V
T
B
1
for all . T = 1. . . . . n, we obtain that

V
T
l
1
}
= 0 for every l. ; = 2. . . . . n. This immediately yields
d
n
g
/}
= 1
n
'1
/
. 1
}
)
= 'V
T
m
1
/
. 1
}
) '1
/
. V
T
m
1
}
) = 0 on ,. l. ;. m = 2. . . . . n.
Moreover, since the 1
s are coordinate vectors for = 1. . . . . n, we obtain

d
n
g
1}
= 1
n
'1
1
. 1
}
)
= 'V
T
m
1
1
. 1
}
) '1
1
. V
T
m
1
}
)
= 'V
T
1
1
n
. 1
}
) '1
1
. V
T
m
1
}
) = 0 on ,. m. ; = 2. . . . . n.
Here we used the fact that V
1
T
1
1
n
= 0 on ,, namely, that V
T
1
1
n
has zero normal
components.
If H = H
n
1
n
is the curvature vector of , (which is normal to the curve), then
we have 'V
T
1
1
n
. 1
1
) = H
n
on ,, so we easily deduce that
(2.1) d
n
g
11
= 1
n
'1
1
. 1
1
) = 2'V
T
1
1
n
. 1
1
) = 2H
n
on ,.
We can also prove that the components 1
1n1}
of the curvature tensor are given by
(2.2) 1
1n1}
=
1
2
d
2
}n
g
11
H
n
H
}
.
Indeed, we have
1
1n1}
= '1(1
1
. 1
}
)1
1
. 1
n
) = 'V
T
1
V
T
j
1
1
. 1
n
) 'V
T
j
V
T
1
1
1
. 1
n
)
= 'V
T
1
V
T
j
1
1
. 1
n
) 1
}
'V
T
1
1
1
. 1
n
) 'V
T
1
1
1
. V
T
j
1
n
)
= 'V
T
1
V
T
j
1
1
. 1
n
) 1
}
'V
T
1
1
1
. 1
n
)
= 'V
T
1
V
T
j
1
1
. 1
n
) 1
}
1
1
'1
1
. 1
n
) 1
}
'1
1
. V
T
1
1
n
)
= 'V
T
1
V
T
j
1
1
. 1
n
) 1
}
'1
1
. V
T
m
1
1
)
= 1
1
'V
T
j
1
1
. 1
n
) 'V
T
j
1
1
. V
T
1
1
n
)
1
2
1
}
1
n
'1
1
. 1
1
)
=
1
2
d
2
}n
g
11

1
4
d
n
g
11
d
}
g
11
.
where here we have used the above properties and the fact that
V
T
j
1
1
= V
T
1
1
}
=
1
2
d
}
g
11
1
1
.
Using (2.1) and (2.2), the above discussion can be summarized in the following
result:
LEMMA 2.1 In the coordinates ( s. .), for . close to 0 the metric coefcients satisfy
g
11
(.) = 1 2
n
nD2
H
n
.
n

1
2
n
n,/D2
(H
n
H
}
1
1n1}
[
,
).
n
.
/
O([.[
3
).
g
1}
(.) =
1
2
n
n,/D2
d
2
n/
g
1}
,
.
n
.
/
O([.[
3
).
g
k}
(.) =
k}

1
2
n
n,/D2
d
2
n/
g
k}
,
.
n
.
/
O([.[
3
).
The second derivatives d
2
n/
g
1}
and d
2
n/
g
k}
can be expressed in terms of the
curvature tensor and the curvature of , through reasoning as for (2.2). However, for
our purposes it is not necessary to have such a formula, so we leave the expansion
of these coefcients in a generic form.
2.2 First and Second Variations of the Length Functional
We next recall the formulas for the variations of the length of a curve with re-
spect to normal displacements. We start with a regular closed curve , in M of
length 1, which we parametrize by arc length, using a parameter s 0. 1|. Then
we consider a two-parameter family of closed curves ,
t
1
,t
2
: 0. 1| M for t
1
. t
2
in a neighborhood of 0 in R such that ,
0,0
,. The length 1(t
1
. t
2
) of ,
t
1
,t
2
is
given by
1(t
1
. t
2
) =
,
t
1
;t
2
Jl =
1
0
' ,
t
1
,t
2
. ,
t
1
,t
2
)
1
2
J s.
where Jl is the arc length parameter and ,
t
1
,t
2
stands for J,
t
1
,t
2
J s. We also
dene the vector elds V and W along ,
t
1
,t
2
as V = d,
t
1
,t
2
dt
1
and W =
d,
t
1
,t
2
dt
2
. In the above coordinates, the vector elds V and W along , can be
written as
V =
n
}D2
V
}
( s)1
}
. W =
n
nD2
W
n
( s)1
n
.
Differentiating 1(t
1
. t
2
) with respect to t
1
we nd
(2.3)
d1(t
1
. t
2
)
dt
1
=
1
0
'V
V
,
t
1
,t
2
. ,
t
1
,t
2
)
' ,
t
1
,t
2
. ,
t
1
,t
2
)
1
2
J s.
Using (2.1), at (t
1
. t
2
) = (0. 0) we have
'V
V
,
t
1
,t
2
. ,
t
1
,t
2
) = V
n
H
n
:
therefore we can write the variation of the length at , in the following way:
(2.4)
d1(t
1
. t
2
)
dt
1
(t
1
,t
2
)D(0,0)
=
1
0
V
n
H
n
J s =
1
0
'V. H)J s.
Using (2.3) we can evaluate the second variation of the length as
d
2
1(t
1
. t
2
)
dt
1
dt
2
=
1
0
'V
W
,
t
1
,t
2
. V
V
,
t
1
,t
2
) ' ,
t
1
,t
2
. V
W
V
V
,
t
1
,t
2
)
' ,
t
1
,t
2
. ,
t
1
,t
2
)
1
2
' ,
t
1
,t
2
. V
V
,
t
1
,t
2
)' ,
t
1
,t
2
. V
W
,
t
1
,t
2
)
' ,
t
1
,t
2
. ,
t
1
,t
2
)
3
2
J s.
so at (t
1
. t
2
) = (0. 0) we nd
d
2
1(t
1
. t
2
)
dt
1
dt
2
(t
1
,t
2
)D(0,0)
=
1
0
'V
W
,. V
V
,) ' ,. V
W
V
V
,) ' ,. V
V
,)' ,. V
W
,)| J s.
By using the denition of the Riemann tensor and the fact that V and W are coor-
dinate vector elds (so that V. W| = 0), the last formula yields
d
2
1(t
1
. t
2
)
dt
1
dt
2
(0,0)
=
1
0
'V
P ,
W. V
P ,
V) ' ,. V
W
V
P ,
V) ' ,. V
P ,
V)' ,. V
P ,
W)
J s
=
1
0
'V
P ,
W. V
P ,
V) '1(W. ,)V. ,) ' ,. V
P ,
V)' ,. V
P ,
W)
J s
1
0
'V
P ,
,. V
W
V)J s.
Here, we have used the fact that
g( ,. V
W
V
P ,
V) = '1(W. ,)V. ,) ,'V
W
V. ,) 'V
P ,
,. V
W
V)
and

1
0
,'V
W
V. ,)J s = 0. Since V
T
l
1
}
= 0 on , for l. ; = 2. . . . . n, we have
V
W
V =
},nD2,...,n
W
n
V
}
V
T
m
1
}
==
1
0
'V
P ,
,. V
W
V)J s = 0.
Moreover, recalling (2.1), we obtain
V
P ,
V =
n
}D2
V
}
1
}

n
}D2
V
}
V
T
1
1
}
=
n
}D2
V
}
1
}

n
}D2
H
}
V
}
1
1
.
This implies, at ,,
'V
P ,
W. V
P ,
V) '1(W. ,)V. ,) ' ,. V
P ,
V )' ,. V
P ,
W) =
n
}D2
V
}

W
}
},/D2
1
1}1/
V
}
W
/
.
In this way the second variation of the length at , becomes
(2.5)
d
2
1(t
1
. t
2
)
dt
1
dt
2
(t
1
,t
2
)D(0,0)
=
1
0
V
}

W
}
},/D2
1
1}1/
V
}
W
/
J s.
2.3 Determining the Phase Factor
In this section we formally derive the asymptotic prole of the solutions to
(NLS
t
) that concentrate near some curve ,, and we determine some necessary
conditions satised by the limit curve. For doing this, using the coordinates ( s. .)
introduced in Subsection 2.1, we look for approximate solutions [( s. .) of (NLS
t
)
making the ansatz
[( s. .) = e
i
f.N s/
"
h( s)U
k( s).
c
. s 0. 1|. . R
n1
.
where the function U is the unique radial solution (see [12, 22, 30, 51]) of the
problem
(2.6)
Û U = U
;
in R
n1
.
U(.) 0 as [.[ o.
U > 0 in R
n1
.
and where the functions } , h, and k are periodic on 0. 1| and have to be deter-
mined. With some easy computations we obtain
d[
d s
=
i}
0
( s)
c
h( s)U
k( s).
c
e
i
f.N s/
"
e
i
f.N s/
"
h
0
( s)U
k( s).
c
e
i
f.N s/
"
h( s)k
0
( s)V
+
U
k( s).
c

.
c
.
d
2
[
d s
2
=
i
}
00
( s)
c
h( s)U
k( s).
c
2i
}
0
( s)
c
h
0
( s)U
k( s).
c
2i
}
0
( s)
c
h( s)k
0
( s)V
+
U
k( s).
c

.
c
(}
0
( s))
2
c
2
h( s)U
k( s).
c
2h
0
( s)k
0
( s)VU
k( s).
c

.
c
h( s)k
00
( s)V
+
U
k( s).
c

.
c
h( s)(k
0
( s))
2
V
2
+
U
k( s).
c
.
c
.
.
c
h
00
( s)U
k( s).
c
e
i
f.N s/
"
.
and also
^
+
[( s. .) =
(k( s))
2
c
2
^
+
U
k( s).
c
e
i
f.N s/
"
h( s).
Since U decays to 0 at innity (exponentially indeed, by the results in [22]), and
since the function [ is scaled of order c near the curve ,, in a rst approximation
we can assume the metric g of M to be at in the coordinates ( s. .); see the
expansions in Lemma 2.1.
We look now at the leading terms in (NLS
t
), which are of order 1. Since ^
v
[
is multiplied by c
2
, we have to focus on the terms of order 1c
2
in the Laplacian
of [. In the above expressions of d[d s, d
2
[d s
2
, and ^
+
[, we have that the
function U and its derivatives are of order 1 when [.[ = O(c); therefore when
the variables . appear as factors in these expressions, we consider them to be of
order c. For example, V
2
U(
k(N x)+
t
)
+
t
.
+
t
| will be regarded as a term of order 1.
With these criteria, using the above computations and assumptions, and if we
impose that the leading terms in (NLS
t
) vanish, we obtain
k
2
( s)h( s)^
+
U
k( s).
c
h( s)
V( s) (}
0
( s))
2
k( s).
c
=
h( s)
;
U
k( s).
c
;
.
From (2.6), we have the two relations
(2.7) k
2
( s) = h( s)
;1
and V( s) (}
0
( s))
2
| = k( s)
2
= h( s)
;1
.
We next obtain an equation for } , which is derived looking at the next-order
expansion of (NLS
t
). The next coefcient arises from the terms of order
1
t
in
^
v
[, which are given by
i
}
00
( s)
c
h( s)U
k( s).
c
2
}
0
( s)
c
h
0
( s)U
k( s).
c
2
}
0
( s)
c
h( s)k
0
( s)V
+
U
k( s).
c

.
c
e
i
f.N s/
"
.
Multiplying this expression by U(
k(N x)+
t
), integrating in . R
n1
, and if we im-
pose that this integral vanishes as well, we get
0 = }
00
( s)h( s)
R
n1
U
2
k( s).
c
J. 2h
0
( s)}
0
( s)
R
n1
U
2
k( s).
c
J.
2}
0
( s)h( s)k
0
( s)
R
n1
U
k( s).
c
V
+
U
k( s).
c

.
c
J..
Integrating by parts and reasoning as for the usual Pohozaevs identity, we obtain
that } must satisfy
}
00
( s)h( s) 2}
0
( s)h
0
( s) (n 1)}
0
( s)h( s)
k
0
( s)
k( s)
= 0.
This is solvable in }
0
( s) and gives, for an arbitrary constant A,
(2.8) }
0
( s) = Ak( s)
n1
h( s)
2
= Ah( s)
.n1/.p1/
2
2
.
p1
V+A h
2
h
h
2
2
h
h
p1
V+A h
2
FIGURE 2.1. The graphs of V A
2
h
2c
and h
;1
for ] <
nC2
n2
and for
] =
nC2
n2
with A < 1.
where we have used equation (2.7) for k. Now we can solve the equation for h( s)
depending on the potential V( s) and the above constant A. In fact, we get that h( s)
should solve
(2.9) V( s) A
2
h( s)
2c
:= V( s) A
2
h( s)
(n1)(;1)4
= h( s)
;1
.
where we have set
(2.10) o =
(n 1)(] 1)
2
2.
Remark 2.2. We notice that, assuming Ato be small enough (depending on V and
]), the above equation is always solvable in h( s). More precisely, when ] <
nC2
n2
(and hence when 2o < ] 1), the solution is also unique. For ] _
nC2
n2
there
might be a second solution. In this case, we just consider the smallest one, which
stays uniformly bounded (both from above and below) when A is small enough;
see Figures 2.1 and 2.2.
Remark 2.3. In the above expansions, considering the terms of order c, as already
noticed, we considered the metric g to be at near the curve ,, and we tacitly as-
sumed the potential V to depend only on the variable s. Indeed, expanding the
Laplace-Beltrami operator and the potential V and taking the variables . into ac-
count, we obtain an extra term of order c that does not affect our computations
since it turns out to be odd in ., so it vanishes once multiplied by U(k( s).c)
and integrated over R
n1
. For more details, we refer to Section 3, where precise
estimates are worked out (in a system of coordinates scaled in c).
2.4 The Euler Equation
Using the heuristic considerations of the previous subsection, we now compute
the energy of an approximate solution [ concentrated near a closed curve ,, and
2 2
h
p1
h
V+A h
2
p1
h
V+A h
2
h
FIGURE 2.2. The graphs of V A
2
h
2c
and h
;1
for ] =
nC2
n2
with
A _ 1 and for ] >
nC2
n2
with Asmall.
then nd the ,s for which this energy is stationary. We let [
,,A
denote the function
constructed in Subsection 2.3. In order for the function [
,,A
to be globally well
dened, we need to impose one more restriction, namely that [
,,A
is periodic in
s with period 1. This is equivalent to requiring that

1
0
}
0
( s)J s be an integer
multiple of 2c, since we have the phase factor e
i(((N x)t)
in the expression of
[
,,A
. From (2.8), we then nd that

1
0
h( s)
c
J s is also an integer multiple of 2c.
Multiplying (NLS
t
) by [
,,A
and integrating by parts, from the fact that [
,,A
is
an approximate solution we nd
1
t
([
,,A
) =
1
2
(c
2
[V
v
[
,,A
[
2
V(.)[[
,,A
[
2
)JV
v

1
] 1
[[
,,A
[
;C1
.
1
2
1
] 1
[[
,,A
[
;C1
JV
v
.
Since [
,,A
is highly concentrated near ,, using the coordinates ( s. .) introduced
in 2.1, we have that
[[
,,A
[
;C1
JV
v
.
1
0
J s
R
n1
h( s)
;C1
k( s).
c
;C1
J..
Using a change of variables, the last two formulas, and (2.7), we nd that
(2.11) 1
t
([
,,A
) .

C
0
c
n1
,
h( s)
0
J s.
where
C
0
=
1
2
1
] 1

R
n1
[U(.)[
;C1
J..
and where we have set
(2.12) 0 = ] 1
1
2
(] 1)(n 1) = ] o 1.
Consider nowa one-parameter family of closed curves ,
t
: 0. 1| M, where t
belongs to a neighborhood of 0 in R and where ,
0
,. We compute next the
approximate value of the derivative in t of the corresponding energy dened by
(2.11).
As in Subsection 2.2 we let V
t
denote the vector eld V
t
( s) =
J,
t
Jt
( s), and we
assume that V := V
0
is normal to ,. For any t near 0, we let k
t
( s), h
t
( s), and
}
t
( s) be dened by (2.7), replacing , by ,
t
and V( s) by V
t
( s) := V(,
t
( s)). Since
we require periodicity of each curve ,
t
in the variable s, we also allow the constant
Agiven in (2.8) to depend on t . Denoting this by A
t
, by the above considerations
we choose A
t
so that the following condition holds for every value of t :
(2.13)
1
0
A
t
h
t
( s)
c
J s =
1
0
}
0
t
( s)J s = const.
Below, we let A
0
t
=
d
dt
A
t
, and we will consider h
t
( s) to be a function of A
t
while V
t
( s) is as implicitly dened in (2.9). From (2.4),
dV
t
( s)
dt
tD0
= 'V
1
V( s). V( s)).
and differentiating (2.13) with respect to t at t = 0, we get
1
0
Aoh
c1
dh
dV
'V
1
V. V)J s A
1
0
h
c
'V. H)J s
AA
0
o
1
0
h
c1
dh
dA
J s A
0
1
0
h
c
J s = 0.
where we have set A
0
= A
0
0
and where V
1
V stands for the component of VV
normal to ,. From this formula we obtain the following expression for A
0
:
(2.14) A
0
= A
1
0
oh
c1 Jh
JV
'V
1
V. V) h
c
'V. H)
J s
1
0
Aoh
c1
Jh
JA
h
c
J s
.
Similarly, computing the derivative of the (approximate) energy with respect to t ,
we nd
J1
t
(u
,
t
,A
t
)
Jt
tD0
=
1
0
0h
01
dh
dV
'V
1
V. V) h
0
'V. H) 0A
0
h
01
dh
dA
J s.
Using (2.14), we deduce that the variation is given by
J1
t
(u
,
t
,A
t
)
Jt
tD0
=
1
0
dh
dV
'V
1
V. V)
0h
01
Aoh
c1
1
0
0h
01 Jh
JA
J s
1
0
Aoh
c1
Jh
JA
h
c
J s
J s
1
0
'V. H)
h
0
Ah
c

1
0
0h
01 Jh
JA
J s
1
0
Aoh
c1
Jh
JA
h
c
J s
J s.
Differentiating (2.9) with respect to Aand V , we get
(2.15)
dh
dA
=
2Ah
2c
(] 1)h
;2
2oA
2
h
2c1
= 2Ah
2c
dh
dV
.
so it follows that
(2.16) Aoh
c1
dh
dA
h
c
=
(] 1)h
;1
(] 1)h
0
2oA
2
h
c
.
Similarly, since 0 = ] o 1 (see (2.10) and (2.12)), we deduce that
h
01
dh
dA
=
2Ah
;Cc2
(] 1)h
;2
2oA
2
h
2c1
.
Therefore we also nd
(2.17)
0h
01 Jh
JA
Aoh
c1
Jh
JA
h
c
=
2A0
] 1
.
Hence from the last formulas the variation of the energy becomes
(2.18)
J1
t
([
,
t
,A
t
)
Jt
tD0
=
1
0
'V
1
V. V)
dh
dV
0h
01
2A
2
o0
] 1
h
c1
J s
1
0
'V. H)
h
0
2A
2
0
] 1
h
c
J s.
Also, from the second equality in (2.15), dividing by h
c
, multiplying by
0
;1
, and
using the identity ] o 2 = 0 1, we obtain
h
c
0
] 1
2oA
2
h
c1
dh
dV
0
] 1
= 0h
;c2
dh
dV
= 0h
01
dh
dV
.
Using (2.18) and the last formula, we get the following simplied expression:
J1
t
([
,
t
,A
t
)
Jt
tD0
=
1
0
0
] 1
h
c
'V
1
V. V) 'V. H)
] 1
0
h
;1
2A
2
h
2c
J s.
Therefore the stationarity condition for the energy (under the constraint (2.13))
becomes 'V
1
V. V) = 'V. H)(
;1
0
h
;1
2A
2
h
2c
) for every normal section V,
namely,
(2.19) V
1
V = H
] 1
0
h
;1
2A
2
h
2c
.
We will see that this formula will be crucial to nding approximate solutions.
Remark 2.4. By (2.16), we have that
d
dA
(Ah
c
) =
(] 1)h
;1
(] 1)h
0
2oA
2
h
c
.
If A is sufciently small (depending on V and ]), then
J
JA
(Ah
c
) > 0. This will
be used in the last section where, for a xed c, we will adjust the value of the
constant Afor obtaining periodicity of the function } .
2.5 Second Variation and Nondegeneracy Condition
We next evaluate the second variation of the Euler functional. As in Subsec-
tion 2.2 we consider a two-parameter family of closed curves ,
t
1
,t
2
, where t
1
and t
2
are two real numbers belonging to a small neighborhood of 0 in R, and
where ,
0,0
= ,. As before, we require the constraint (2.13) along the whole two-
dimensional family of curves, and we assume that the functions }. h, and k and the
constant A depend on t
1
and t
2
, and we will use the notation A
t
1
,t
2
, etc. Keeping
this in mind, we dene the two vector elds
V
t
1
,t
2
=
d,
t
1
,t
2
dt
1
. W
t
1
,t
2
=
d,
t
1
,t
2
dt
2
.
and we can assume that V := V
0,0
and W := W
0,0
are normal to the initial
curve ,.
With some computations, which are worked out in [36, sec. 5.2], we nd that, at
(t
1
. t
2
) = (0. 0),
d
2
1
t
(u
v
t
1
;t
2
,A
t
1
;t
2
)
dt
1
dt
2
(2.20)
=
1
0
h
0
2A
2
0
] 1
h
c
V
}

W
}
},n
1
1}1n
V
}
W
n
J s
0
] 1
1
0
((V
1
)
2
V )V. W| 'V
1
V. V)'H. W)
'V
1
V. W)'H. V)
h
c
J s
o0
] 1
1
0
h
c1
dh
dV
'V
1
V. V)'V
1
V. W)J s
A
0
1
A
0
2
20
] 1
1
0
Aoh
c1
dh
dA
h
c
J s.
Here
A
0
/
=
dA
ot
1
,t
2
dt
/
(t
1
,t
2
)D(0,0)
.

V
}
=
JV
}
J s
. and

W
}
=
JW
}
J s
.
where the V
}
and W
}
are the components of V and W with respect to the basis
(1
}
)
}
introduced in Subsection 2.1.
Integrating by parts and using (2.14), from the last formula we derive that the
nondegeneracy condition is equivalent to the invertibility of the linear operator
J : ,(N,) ,(N,) (from the family of smooth sections of the normal bundle to
, into itself) whose components are dened by
(JV)
n
=
h
0
2A
2
0
] 1
h
c

V
n
}
1
1}1n
V
}
h
01
2A
2
o
] 1
h
c1
h
0

V
n
0
] 1
h
c
((V
1
)
2
V )(V. 1
n
) H
n
'V
1
V. V)
'H. V)'V
1
V. 1
n
)
o0
] 1
h
(cC1)
dh
dV
'V
1
V. V)'V
1
V. 1
n
)
20
] 1
AA
0
1
oh
c1
dh
dV
'V
1
V. 1
n
) h
c
H
n
.
(2.21)
where h
0
=
dh(N x)
d N x
.
Using (2.19) and some other elementary computations (see subsection 2.5 in
[36]), we also nd
(JV)
n
=
h
0
2A
2
0
] 1
h
c

V
n
0
h
01
2A
2
o
] 1
h
c1
h
0

V
n
0
] 1
h
c
((V
1
)
2
V )V. 1
n
|
1
2
h
0
2A
2
0
] 1
h
c
}
(d
2
}n
g
11
)V
}
2AA
0
1
(0 o)h
;1
(] 1)h
0
2oA
2
h
c
|
H
n
H
n
'H. V)
(] 1)(3
c
0
)h
20
16c0A
4
;1
h
2c
2A
2
(5o 30)h
0Cc
(] 1)h
0
2A
2
oh
c
.
(2.22)
The latter expression of JV is going to be useful later on. We summarize the results
of this section in the following proposition.
PROPOSITION 2.5 Consider the functional on curves

,
h
0
( s)J s restricted to the
set I in (1.13). Then the stationarity condition is (2.19), and the nondegeneracy of
a critical point is equivalent to the invertibility of the operator J in (2.22).
3 Some Preliminary Expansions
In this section we nd a family of approximate solutions to the scaled equation
(1.16). We consider a simple closed curve , that is stationary within the class I,
namely satisfying (2.19). First, we introduce some convenient coordinates near the
scaled curve ,
t
=
1
t
,, expanding the Laplace-Beltrami operator with respect to
the scaled metric in powers of c. Then, using these expansions, we construct the
approximate solutions formally solving (1.16) up to order c.
3.1 Choice of Coordinates in M
"
and Expansion of the Metric Coefcients
Using the coordinates ( s. .) of Section 2 dened near ,, for some smooth nor-
mal section ( s) in N,, we dene the following new coordinates (s. z) (here and
below we use the notation s = cs) near
1
t
,
(3.1) z = . (cs). z R
n1
.
In this choice we are motivated by the fact that in general we allow the approximate
solutions to be tilted normally to ,
t
, where the tilting depends (slowly) on the
variable s; this allows some extra exibility in the construction, as in [19, 34,
39]. As we will see, the choice of is irrelevant for solving (1.16) up to order
c; on the other hand, the nondegeneracy assumption will be necessary to guarantee
solvability of the equation up to higher orders.
We denote by g
B
the metric coefcients in the new coordinates (s. z). Since
. = z (cs), it follows that
g
CT
=
B
g
B
d.
dz
C
d.
B
dz
T
.
Explicitly, we then nd
g
11
= g
11
:C
2c
0
}
g
1}
:C
c
2
},n
0
}
(cs)
0
n
(cs)g
}n
:C
.
g
1}
= g
1}
:C
c
0
n
(cs)g
}n
:C
. g
}n
= g
}n
[
:C
.
At this point, it is convenient to introduce some notation. For a positive inte-
ger q, we denote by 1
q
(z), 1
q
(z. ), and 1
q
(z. .
0
) error terms that satisfy,
respectively, the following bounds, for some positive constants C and J:
[1
q
(z)[ _ Cc
q
(1 [z[
d
).
[1
q
(z. )[ _ Cc
q
(1 [z[
d
).
[1
q
(z. ) 1
q
(z.

)[ _ Cc
q
(1 [z[
d
) [

[|.
and
[1
q
(z. .
0
)[ _ Cc
q
(1 [z[
d
).
[1
q
(z. .
0
) 1
q
(z.

.

0
)[ _ Cc
q
(1 [z[
d
)[

[ [
0

0
[|.
We also introduce error terms involving second derivatives of , 1
q
(z. .
0
.
00
), which satisfy
[1
q
(z. .
0
.
00
)[ _ Cc
q
(1 [z[
d
) Cc
qC1
(1 [z[
d
)[
00
[
and
[1
q
(z. .
0
.
00
) 1
q
(z.

.

0
.

00
)[
_ Cc
q
(1 [z[
d
)[

[ [
0

0
[|(1 c([
00
[ [
00
[))
Cc
qC1
(1 [z[
d
)[
00

00
[.
Using the expansion of the metric coefcients g
B
in Lemma 2.1 and this notation,
we then obtain
g
11
= 1 2c
n
nD2
H
n
(z
n

n
)
1
2
c
2
n
n,/D2
d
2
n/
g
11
,
(z
n

n
)(z
/

/
)
c
2
[
0
[
2
1
3
(z. .
0
).
(3.2)
g
1}
= c
0
}

1
2
c
2
n
n,/D2
d
2
n/
g
1}
,
(z
n

n
)(z
/

/
) 1
3
(z. .
0
).
g
k}
=
k}

1
2
c
2
n
n,/D2
d
2
n/
g
k}
,
(z
n

n
)(z
/

/
) 1
3
(z. .
0
).
(3.3)
Next we compute the inverse metric coefcients. Recall that, given a formal
expansion of a matrix as M = 1 c c
2
T, we have
M
1
= 1 c c
2
2
c
2
T.
In our specic case the matrix is the following:
(3.4) =
2
n
nD2
H
n
(z
n

n
)
0
}
0
}
0
.
and the elements of its square are given by
(
2
)
11
= 4
nD2
H
n
(z
n

n
)
}
(
0
}
)
2
.
(
2
)
1}
= 2
nD2
H
n
(z
n

n
)
(
0
}
). (
2
)
/}
= (
0
/
)(
0
}
).
Therefore, using the above formula, the inverse coefcients are
g
11
= 1 2c
n
nD2
H
n
(z
n

n
)
1
2
c
2
n
n,/D2
d
2
n/
g
11
,
(z
n

n
)(z
/

/
)
4c
2
nD2
H
n
(z
n

n
)
2
1
3
(z. .
0
).
We also get
g
1}
= c
0
}

1
2
c
2
n
n,/D2
d
2
n/
g
1}
,
(z
n

n
)(z
/

/
)
2c
2
nD2
H
n
(z
n

n
)
0
}
1
3
(z. .
0
).
Moreover,
d
}
( g
1}
) = c
2
n
/D2
d
2
/}
g
1}
,
(z
/

/
) 2c
2
H
}
0
}
1
3
(z. .
0
).
Similarly, with some simple calculations we also nd
d
1
( g
11
) = 2c
2
n
nD2
(H
n
)
0
(z
n

n
)
2c
2
n
nD2
H
n
0
n
1
3
(z. .
0
.
00
).
Differentiating now g
1}
with respect to the rst variable, we obtain
d
1
( g
1}
) = c
2
00
}
2c
3
nD2
H
n
(z
n

n
)
00
}
1
3
(z. .
0
.
00
).
Analogously, we get
g
k}
=
k}

1
2
c
2
n
n,/D2
d
2
n/
g
k}
,
(z
n

n
)(z
/

/
)
c
2
0
k
0
}
1
3
(z. .
0
).
d
k
( g
k}
) = c
2
n
/D2
d
2
k/
g
k}
,
(z
/

/
) 1
3
(z. .
0
).
Finally, using the formal expansion g
CT
=
CT
c
CT
c
2
T
CT
o(c
2
)
and carefully analyzing the error terms, we obtain
det g = 1
1
2
c tr() c
2
1
8
(tr())
2
1
4
tr(
2
)
1
2
c
2
tr(T) O(c
3
).
From the above expressions in (3.2) and (3.3) we deduce that
det g = 1 c
n
H
n
(z
n

n
)
c
2
1
4
n,/
d
2
n/
g
11
(z
n

n
)(z
/

/
)
1
2
nD2
H
n
(z
n

n
)
1
3
(z. .
0
).
d
n
det g = cH
n
c
2
1
2
/
d
2
n/
g
11
(z
/

/
)
H
n
/
H
/
(z
/

/
)
1
3
(z. .
0
).
Moreover,
d
1
det g = c
2
n
(H
n
)
0
(z
n

n
) c
2
n
H
n
0
n
1
3
(z. .
0
.
00
).
The Laplacian of a smooth function u in coordinates (s. z) has the following
expression:
^
Q v
u =
,B
g
B
d
2
B
u
,B
d
B
( g
B
)d
det g
,B
g
B
d
B
det g
u.
We are going to expand next each of these terms. First, we consider the determinant
of g. Recall that for a matrix of the form 1 c c
2
T, the square root of the
determinant admits the formal expansion
(3.5)

det g = 1
c
2
tr c
2
1
8
(tr )
2
1
4
tr(
2
)
1
2
tr T
o(c
2
).
LEMMA 3.1 Let u be a smooth function. Then in the above coordinates (s. z), we
have that
^
Q v
u = d
2
xx
u ^
:
u c
}
H
}
d
}
u 2c
0
}
d
2
x}
u 2c'H. z )d
2
xx
u
c
2
'H. z )
n,}
H
}
d
}
u
1
2
c
2
d
2
n/
g
11
(z
n

n
)(z
/

/
)d
2
xx
u
4c
2
'H. z )
2
d
2
xx
u c
2
d
2
n/
g
1}
(z
n

n
)(z
/

/
)d
2
x}
u
4c
2
'H. z )
0
}
d
2
x}
u c
2
t,}
0
t
0
}
d
2
t}
u
1
2
c
2
n,/
d
2
n/
g
t}
(z
n

n
)(z
/

/
)d
2
t}
u c
2
'H
0
. z )d
x
u
c
2
/,}
d
2
/}
g
1}
(z
/

/
)d
x
u c
2
00
}
d
}
u
c
2

t,},/
d
2
t/
g
t}
(z
/

/
)d
}
u 2c
3
'H. z )
00
}
d
}
u
1
3
(z. .
0
)d
2
xx
u 1
3
(z. .
0
)d
2
x}
u
1
3
(z. .
0
)d
2
/}
u 1
3
(z. .
0
.
00
)(d
x
u d
}
u).
Moreover, given two smooth normal sections and

and dening the corre-
sponding coordinates
(s. . (cs)) and (s. .

(cs)).
we set u
(s. .) := u(s. . (cs)), u

Q
(s. .) := u(s. .

(cs)). We then have
^
Q v
u
^
Q v
u
Q
= 2c
}
(
0
}

0
}
)d
2
x}
u 2c'H.

)d
2
xx
u c
2
t,}
(
0
t
0
}

0
t

0
}
)d
2
t}
u
1
2
c
2
n,/
d
2
n/
g
t}
2z
n
(
/

/
)
/
(
n

n
)

/
(
n

n
)
d
2
t}
u
c
2
n,/
d
2
n/
g
1}
2z
n
(
/

/
)
/
(
n

n
)

/
(
n

n
)
d
2
x}
u
1
2
c
2
n,/
d
2
n/
g
11
2z
n
(
/

/
)
/
(
n

n
)

/
(
n

n
)
d
2
xx
u
2c
2
/
H
}
z
/
(
0
/

0
/
)
/
(
0
/

0
/
)

0
/
(
/

/
)
d
2
x}
u
4c
2
n,/
H
n
H
/
2z
n
(
/

/
)
/
(
n

n
)

/
(
n

n
)
d
2
xx
u
c
2
}
(
00
}

00
}
)d
}
u c
2
t,},/
d
2
t/
g
t}
(
/

/
)d
}
u c
2
'H.

)
}
H
}
d
}
u
c
2
'H
0
.

)d
x
u c
2
/,}
d
2
/}
g
1}
(
/

/
)d
x
u
2c
3
n}
H
n
(z
n

n
)(
00
}

00
}
)

00
}
(
n

n
)
d
}
u
O(1 [z[
d
)
c
4
([

[ [
0

0
[)[d
2
xx
u[
c
3
([

[ [
0

0
[)([d
2
x}
u[ [d
2
/}
u)[
O(1 [z[
d
)
c
3
([

[ [
0

0
[)
c
4
([
00
[[

[ [
00

00
[)
([d
x
u[ [d
}
u[).
PROOF: The proof is based on the Taylor expansion of the metric coefcients
given above. We recall that the Laplace-Beltrami operator is given by
^
Q v
=
,B
1
det g
d
det g(g
t
)
B
d
B
).
where indices and T run between 1 and n. We can also write
^
Q v
=
,B
g
B
d
2
B
(d
g
B
)d
B

1
det g
g
B
(d
B
det g)d
.
Using the expansion of the metric coefcients determined above and (3.5), we can
easily prove that
B
g
B
d
2
B
u
= ^
:
u d
2
xx
u 2c
0
}
d
2
x}
u 2c'H. z )d
2
xx
u
c
2
/,}
0
/
0
}
d
2
/}
u 4c
2
'H. z )
2
d
2
xx
u
1
2
c
2
n,/
d
2
n/
g
k}
(z
n

n
)(z
/

/
)d
2
k}
u
1
2
c
2
n,/
d
2
n/
g
1}
(z
n

n
)(z
/

/
)d
2
x}
u
4c
2
'H. z )
0
}
d
2
x}
u
1
2
c
2
n,/
d
2
n/
g
11
(z
n

n
)(z
/

/
)d
2
xx
u
1
3
(z. .
0
)(d
2
xx
u d
2
x}
u d
2
/}
u).
,B
d
g
B
d
B
u
= c
2
00
}
d
}
u c
2

i,},/
d
2
k/
g
k}
(z
/

/
)d
}
u
2c
3
'H. z )
00
}
d
}
u 2c
2
'H
0
. z )d
x
u
c
2
/,}
d
2
/}
g
1}
(z
/

/
)d
x
u 1
3
(z. .
0
.
00
)(d
x
u d
}
u).
,B
1
det g
g
B
d
B
det g
u
= c
}
H
}
d
}
u c
2
'H. z )
}
H
}
d
}
u c
2
'H
0
. z )d
x
u
1
2
}
d
2
/}
g
11
(z
/

/
)d
}
u 1
3
(z. .
0
)(d
x
u d
}
u).
The result then follows by collecting these three terms.
3.2 Expansion at First Order in "
In this subsection we solve equation (1.16) up to order c, discarding the terms
that turn out to be of order c
2
and higher. Here and in the next two subsections we
will display formal expansions only: we will assume that all the data are smooth
and write O(c
k
) for terms that appear at the k
th
-order in a formal expansion. Since
all the functions we are dealing with decay exponentially in z, the error terms do
also: precise statements are given in Lemma 3.2 and Proposition 3.3 below.
For the approximate solution as in (1.10), we make a more precise ansatz of the
following form:
(3.6)
[
1,t
(s. z) = e
i
z
f
0
."s/
"
{h(cs)U(k(cs)z) cn
i
in
i
|].
s 0. 2|. . R
n1
.
where

}
0
(cs) = }(cs) c}
1
(cs). By direct computation, the rst and second
derivatives of [
1,t
satisfy
d
x
[
1,t
= e
i
z
f
0
."s/
"
i

}
0
0
(cs)h(cs)U(k(cs)z) ch
0
(cs)U(k(cs)z)
ch(cs)k
0
(cs)VU(k(cs)z) z
e
i
z
f
0
."s/
"
i c}
0
n
i
c}
0
n
i
| O(c
2
).
d
i
[
1,t
= e
i
z
f
0
."s/
"
h(cs)k(cs)d
i
U(k(cs)z) cd
i
n
i
i cd
i
n
i
|.
d
2
xx
[
1,t
= (

}
0
0
)
2
hU(kz)e
i
z
f
0
."s/
"
i ce
i
z
f
0
."s/
"
}
00
hU(kz) 2}
0
h
0
U(kz) 2}
0
hk
0
VU(kz) z|
c}
02
e
i
z
f
0
."s/
"
n
i
in
i
| O(c
2
).
d
2
/}
[
1,t
= e
i
z
f
0
."s/
"
h(cs)k
2
(cs)d
2
/}
U(k(cs)z) cd
2
/}
n
i
i cd
2
/}
n
i
.
d
2
x}
[
1,t
= e
i
z
f
0
."s/
"
i

}
0
0
(cs)h(cs) ch
0
(cs)
k(cs)d
}
U(kz)
ce
i
z
f
0
."s/
"
h(cs)k
0
(cs)
/
d
2
/}
U(kz)z
/
d
}
U(kz)
i c}
0
(cs)d
}
n
i
(cs. z)e
i
z
f
0
."s/
"
c}
0
(cs)d
}
n
i
(cs. z)e
i
z
f
0
."s/
"
O(c
2
).
Similarly, the potential V satises
V(c.) = V(cs) c'V
1
V. z ) O(c
2
).
Expanding (1.16) in powers of c, we obtain
e
i
z
f
0
."s/
"
(^
v
[
1,t
V(c.)[
1,t
[[
1,t
[
;1
[
1,t
) = cR
i
i cR
i
O(c
2
).
with
R
i
= L
i
n
i
2}
0
}
0
1
hU 2}
02
hU(kz)'H. z ) hk'H. VU(kz))
'V
1
V. z )hU(kz).
(3.7)
R
i
= L
i
n
i
}
00
hU(kz) 2}
0
h
0
U(kz) 2}
0
hk
0
VU(kz) z|
2
0
}
}
0
hkd
}
U(kz)|.
(3.8)
and where we have dened the two operators L
i
and L
i
as
L
i
n = ^
:
n (V }
02
)n ]h
;1
U(kz)
;1
n.
L
i
n = ^
:
n (V }
02
)n h
;1
U(kz)
;1
n.
It is well-known (see, e.g., [48]) that the kernel of L
i
is generated by the (n 1)-
tuple of functions d
2
U(k). . . . . d
n
U(k), while that of L
i
is one-dimensional and
generated by U(k).
We choose the functions n
i
and n
i
in such a way that R
i
and R
i
vanish. Since
L
i
is Fredholm, the solvability condition for n
i
is that the right-hand side of this
equation is orthogonal in 1
2
(R
n1
) to d
2
U(k). . . . . d
n
U(k). Therefore, to get
solvability, we should multiply the right-hand side by each of these functions and
get 0. The same holds true for n
i
, but by replacing the functions d
:
j
U(k) by
U(k).
We discuss the solvability in n
i
rst. Writing this equation as L
i
n
i
= f, we
can multiply it by U(k) and use the self-adjointness of L
i
to get
0 =
R
n1
n
i
L
i
U(k) =
R
n1
U(k)L
i
n
i
=
R
n1
fU(k).
Following the computations of Subsection 2.3, this condition yields
}
00
hk
(n1)
2}
0
h
0
k
(n1)
= (n 1)hk
0
k
n
}
0
.
which implies
}
0
= A
k
n1
h
2
= Ah
c
.
This equation is nothing but (2.8), and hence the solvability is guaranteed. Since
L
i
clearly preserves the parity in z, we can decompose n
i
in its even and odd
parts as
n
i
= n
i,e
n
i,o
.
with n
i,e
and n
i,o
solving, respectively, the equations
L
i
n
i,e
= }
00
hU(kz) 2}
0
h
0
U(kz) 2}
0
hk
0
VU(kz) z|.
L
i
n
i,o
= 2
0
}
}
0
hkd
}
U(kz)|.
where the right-hand sides are, respectively, the even and odd parts of the datum
in (3.8). We notice that, since the kernel of L
i
consists of even functions, only the
even part of the equation plays a role in the solvability, since the product with the
odd part vanishes by oddness.
Indeed, (3.7) and (3.8) can be solved explicitly, and the solutions are given by
(3.9) n
i,e
=
] 1
4
}
0
h
0
[z[
2
U(kz). n
i,o
=
0
}
}
0
hz
}
U(kz).
In fact, as we can easily check, we have the following relations:
L
i
(z
}
U(kz)) = 2kd
}
U(kz).
L
i
([z[
2
U(kz)) = 2(n 1)U(kz) 4kVU(kz) z.
which imply the above claim (here we also used (2.7) and some manipulations).
Turning to n
i
, if we multiply by d
}
U, integrate by parts, and use some scaling,
we nd that the following condition holds true, for ; = 2. . . . . n:
2H
}
(}
0
)
2
R
n1
U
2
Jz
k
2
n 1
R
n1
[VU[
2
Jz
'V
1
V. 1
}
)
R
n1
U
2
Jz = 0.
Using (2.7), we get equivalently, for ; = 2. . . . . n,
2H
}
A
2
h
2c
R
n1
U
2
Jz
h
;1
n 1
R
n1
[VU[
2
Jz
'V
1
V. 1
}
)
R
n1
U
2
Jz = 0.
From a Pohozaev-type identity (playing with (2.6) and integrating by parts) we nd
R
n1
[VU(z)[
2
Jz =
(n 1)(] 1)
(3 n)(] 1) 2(n 1)
R
n1
U(z)
2
Jz (3.10)
=
(n 1)(] 1)
20
R
n1
U(z)
2
Jz.
By using this formula, the solvability condition then becomes
H
}
(] 1)
h
;1
0
2A
2
h
2c
= 'V
1
V. 1
}
). ; = 2. . . . . n.
which is nothing but the stationary condition (2.19). Therefore, since we are indeed
assuming this condition, the solvability for n
i
is also guaranteed. As for n
i
, we
can decompose n
i
in its even and odd parts as
n
i
= n
i,e
n
i,o
.
where n
i,e
and n
i,o
solve, respectively,
L
i
n
i,e
= 2}
0
}
0
1
hU 2(}
0
)
2
hU(kz)'H. ) 'V
1
V. )hU(kz). (3.11)
L
i
n
i,o
= 2(}
0
)
2
hU(kz)'H. z) hk
}
H
}
d
}
U(kz)
'V
1
V. z)hU(kz).
(3.12)
Using the Euler equation, we get
L
i
n
i,o
= h
}
H
}
kd
}
U h
;1
] 1
0
z
}
U
.
It is also convenient to have the explicit expression of n
i
. We notice rst that
L
i
1
(] 1)h
;1
U(kz)
1
2k
VU(kz) z
= U(kz).
Hence it follows
n
i,e
=
h'V
1
V 2(}
0
)
2
H. ) 2}
0
}
0
1
h
1
(] 1)h
;1
U(kz)
1
2k
VU(kz) z
.
(3.13)
Using (2.19) we nally nd
n
i,e
=
] 1
0
h
;
'H. ) 2}
0
}
0
1
h
1
(] 1)h
;1
U(kz)
1
2k
VU(kz) z
.
By the above computations (and the comments at the beginning of this subsection)
we obtain the following result:
LEMMA 3.2 Suppose h( s) and }( s) satisfy (1.11) and (1.12) for some A > 0;
assume also that the curve , satises (1.15). Then there exist two smooth functions
n
i
( s. z) and n
i
( s. z) for which the terms R
i
and R
i
in (3.7)(3.8) vanish iden-
tically. Therefore, the function [
1,t
in (3.6) satises (1.16) up to an error of the
form 1
2
(z)e
k(tx)j:j
.
3.3 Expansions at Second Order in "
Next we compute the terms of order c
2
in the above expression. Adding a cor-
rection c
2
i
i
i
| to the function in (3.6), we dene an approximate solution of
the form
(3.14) [
2,t
(s. z) = e
i
z
f
0
."s/
"
{h(cs)U(k(cs)z) cn
i
in
i
| c
2
i
i
i
|]
with s 0. 2| and . R
n1
, where

}
0
= }(cs) c}
1
(cs). The rst and second
derivatives of [
2,t
are given by
e
i
Q
f
0
."s/
"
d
x
[
2,t
=

i

}
0
0
(cs)h(cs)U(k(cs)z) ch
0
(cs)U(k(cs)z)
ch(cs)k
0
(cs)VU(k(cs)z) z

i c

}
0
0
n
i
c

}
0
0
n
i

i c
2

}
0
0
i
c
2

}
0
0
c
2
(d
x
n
i
in
i
) O(c
3
).
e
i
Q
f
0
."s/
"
d
}
[
2,t
=

h(cs)k(cs)d
}
U(k(cs)z) cd
}
n
i
i cd
}
n
i
c
2
d
}
i
i c
2
d
}
.
e
i
Q
f
0
."s/
"
d
2
[
2,t
ds
2
= (

}
0
0
)
2
hU(kz) i c

}
00
0
hU(kz) 2

}
0
0
h
0
U(kz) 2

}
0
0
hk
0
VU(kz) z|
c

}
02
0
n
i
in
i
| c
2

}
02
0

i
i
i
|
c
2
2

}
0
0
d
x
n
i
h
00
U(kz) 2h
0
k
0
VU z hk
00
VU z
hk
02
V
2
U(kz)z. z|

}
00
0
n
i
i c
2
2

}
0
0
d
x
n
i

}
00
0
n
i
| O(c
3
).
e
i
Q
f
0
."s/
"
d
2
/}
[
1,t
=

h(cs)k
2
(cs)d
2
/}
U(k(cs)z) cd
2
/}
n
i
i cd
2
/}
n
i
c
2
d
2
/}
i
i c
2
d
2
/}
.
e
i
Q
f
0
."s/
"
d
2
x}
[
2,t
=

i

}
0
0
(cs)h(cs)k(cs) ch
0
(cs)k(cs) ch(cs)k
0
(cs
d
}
U(kz)
ch(cs)k(cs)k
0
(cs)
/
d
2
/}
U(kz)z
/
d
}
U(kz) i c

}
0
0
d
}
n
i
c

}
0
0
(cs)d
}
n
i
i c
2

}
0
0
d
}
i
c
2

}
0
0
(cs)d
}
i
c
2
d
2
x}
n
i
i c
2
d
2
x}
n
i
.
We also have the formal expansion
e
i
Q
f
0
."s/
"
[[
2,t
[
;1
[
2,t
= h
;
[U[
;1
U ]ch
;1
[U[
;1
n
i
i ch
;1
[U[
;1
n
i
1
2
](] 1)c
2
h
;2
[U[
;3
Un
2
i
1
2
(] 1)c
2
h
;2
[U[
;3
Un
2
i
i(] 1)c
2
h
;2
[U[
;3
Un
i
n
i
]c
2
h
;1
[U[
;1
i
i c
2
h
;1
[U[
;1
i
O(c
3
).
Similarly, expanding V up to order c
2
, we have
V(c.) = V(cs) c'V
1
V. z )
1
2
c
2
(V
1
)
2
V z . z | 1
3
(z. ).
Using the expansions of Subsection 2.3, we obtain
e
i
Q
f
0
."s/
"
(^
v
[
2,t
V(c.)[
2,t
[[
2,t
[
;1
[
2,t
)
= c
2
(

R
i
i

R
i
)
= c
2
(

1
i,e

1
i,o
) c
2
i(

1
i,e

1
i,o
)
c
2
(

1
i,e,(
1

1
i,o,(
1
) c
2
i(

1
i,e,(
1

1
i,o,(
1
)
c
2
L
i
i
c
2
i L
i
i
O(c
3
).
where
1
i,e
=
1
2
(}
0
)
2
hU(kz)
/,n
d
2
/n
g
11
(z
n
z
/

n
/
) (3.15)
2(}
0
)
2
'H. n
i,e
n
i,o
z)
4(}
0
)
2
hU(kz)'H. z)
2
'H. )
2
| 2}
0
d
x
n
i,e
}
00
n
i,e

h
00
U(kz) 2h
0
k
0
VU(kz) z hk
00
VU(kz) z
h(k
0
)
2
V
2
U(kz)z. z|
2
0
}
}
0
d
}
n
i,o
1
2
/,n
d
2
/n
g
t}
(z
n
z
/

n
/
)
0
t
0
}
hk
2
d
2
t}
U(kz)
hk
/,n,}
d
2
/n
g
n}
z
/
d
}
U(kz)
n
H
n
d
n
n
i,o

hk'H. z)
n
H
n
d
n
U(kz)
kh
'H. z)H
n
1
2
/
d
2
n/
g
11
z
/
d
n
U(kz)
1
2
](] 1)h
;2
U(kz)
;2
(n
2
i,e
n
2
i,o
)
1
2
(] 1)h
;2
U(kz)
;2
(n
2
i,e
n
2
i,o
)
'V
1
V. n
i,o
z n
i,e
)
1
2
n,}
d
2
n}
V(z
n
z
}

n
}
)hU(kz).
(3.16)
1
i,o
= (}
0
)
2
hU(kz)
/,n
d
2
/n
g
11
z
n
/
8(}
0
)
2
hU(kz)'H. z)'H. )
2(}
0
)
2
'H. n
i,e
z n
i,o
) 2}
0
d
x
n
i,o
}
00
n
i,o
2h
0
k
0
}
d
}
U(kz)
2h
0
k
0
}
/
d
2
/}
U(kz)z
/
d
}
U(kz)
2}
0
0
}
d
}
n
i,e
hk'H. )
n
H
n
d
n
U(kz)
1
2
/,n
d
2
/n
g
t}
(z
n
/
z
/
n
)
hk
2
d
2
t}
U(kz)
hk
00
}
d
}
U(kz)
n
H
n
d
n
n
i,e
hk
'H. )H
n
1
2
/
d
2
n/
g
11
d
n
U(kz)
](] 1)h
;2
U(kz)
;2
n
i,e
n
i,o
(] 1)h
;2
U(kz)
;2
n
i,e
n
i,o
'V
1
V. n
i,e
z n
i,o
)
},/
d
2
}/
Vz
}
/
hU(kz)

},/,n
d
2
/n
g
n}
hkd
}
U(kz).
1
i,e
= 2
}
00
hU(kz) 2}
0
h
0
U(kz) 2}
0
hk
0
VU(kz) z
'H. )
2(}
0
)
2
'H. n
i,e
n
i,o
z) 2}
0
d
x
n
i,e
}
00
n
i,e
2}
0
0
}
d
}
n
i,o
2}
0
hk
}
d
}
U(kz)
2'H. z)
0
}
1
2
/,n
d
2
/n
g
1}
(z
n
/
z
/
n
)
}
0
hU(kz)
n
(d
2
1n
g
11
n
2'H.
0
))
}
0
h
2'H.
0
)
},/
d
2
/}
g
1}
U(kz)
1
2
}
0
h
/
d
1/
g
11
U(kz)
}
H
}
d
}
n
i,o
(] 1)h
;2
U(kz)
;2
(n
i,e
n
i,e
n
i,o
n
i,o
)
'V
1
V. n
i,o
z n
i,e
).
(3.17)
1
i,o
= 2
}
00
hU(kz) 2}
0
h
0
U(kz) 2}
0
hk
0
VU(kz) z
'H. z) (3.18)
i
H
}
d
}
n
i,e
2(}
0
)
2
'H. n
i,e
z n
i,o
) 2}
0
d
x
n
i,o
}
00
n
i,o
2}
0
0
}
d
}
n
i,e
2}
0
hk
}
d
}
U(kz)
2'H. )
0
}
1
2
/,n
d
2
/n
g
1}
(z
n
z
/

/
n
)
}
0
hU(kz)
n
d
2
1n
g
11
z
n
}
0
h
},/
d
2
/}
g
1}
z
/
U(kz)
1
2
}
0
h
/
d
2
1/
g
11
z
/
U(kz)
(] 1)h
;2
U(kz)
;2
(n
i,e
n
i,o
n
i,o
n
i,e
)
'V
1
V. n
i,e
z n
i,o
).
We used the notation

1
i,e,(
1
,

1
i,o,(
1
,

1
i,e,(
1
, and

1
i,o,(
1
for the terms involving
}
1
, namely,
(3.19)
1
i,e,(
1
= (}
0
1
)
2
hU 2}
0
}
0
1
n
i,e
4'H. )}
0
}
0
1
hU
2](] 1)h
;2
[U[
;2
h
2
}
02
}
02
1

U
2
2}
0
}
0
1
h'V
1
V. )

U 4'H. )(}
0
)
3
}
0
1
h

U
2]
(] 1)
2
0
h
2;1
}
0
}
0
1
'H. )U
;2
U
2
.
(3.20)
1
i,o,(
1
= 2}
0
}
0
1
n
i,o
4'H. z)}
0
}
0
1
hU
2](] 1)}
0
}
0
1
h
;1
U
;2
Un
i,o
2}
0
}
0
1
h'V
1
V. z)

U 2H
}
}
0
}
0
1
hkd
}

U
4'H. z)(}
0
)
2
}
0
}
0
1
h

U.
(3.21)
1
i,e,(
1
= 2h
0
}
0
1
U 2h}
0
1
k
0
VU z 2}
0
}
0
1
n
i,e
}
00
1
hU
4}
0
d
x
(h}
0
}
0
1

U) 2}
00
h}
0
}
0
1

U
2(] 1)h
;1
[U[
;2
}
0
}
0
1

Un
i,e
.
(3.22)
1
i,o,(
1
= 2}
0
}
0
1
n
i,o
2(] 1)h
;1
[U[
;2
}
0
}
0
1

Un
i,o
4(}
0
)
2
hk}
0
1
0
}
d
}

U 2hk}
0
1
0
}
d
}
U.
where we wrote for brevity
U =
1
(] 1)h
;1
U(kz)
1
2k
VU(kz) z.
Again, we collect the results of this section in one proposition.
PROPOSITION 3.3 Suppose and }
1
are smooth functions on 0. 1|. Let z be the
normal coordinates given in (3.1), and n
i
and n
i
be as in Lemma 3.2. Then, if
[
2,t
is as in (3.14), in the coordinates (s. z) we have
(3.23)
^
Q v
"
[
2,t
V(c.)[
2,t
[[
2,t
[
;1
[
2,t
= c
2
(L
i
i
i L
i
i
)
c
2

1
i,e

1
i,o
i

1
i,e
i

1
i,o

1
i,e,(
1

1
i,o,(
1
i

1
i,e,(
1
i

1
i,o,(
1
1
3
(z)e
k(tx)j:j
.
where the above error terms are given, respectively, in (3.15)(3.22).
It is useful next to evaluate the projections of the errors on the kernels of the
operators L
i
and L
i
to see what the effect of and }
1
is. Concerning L
i
, the
kernel is spanned by i U(k(cs)). To compute this projection, for parity reasons,
we need to multiply

1
i,e
and

1
i,e,(
1
by U(k(cs)) and to integrate over R
n1
.
For brevity, we only display the nal results, referring to [36, sec. 4] for complete
proofs.
Contribution of

1
i,e
:
R
n1
1
i,e
U(kz) =
2A
h
] 1
20
1

R
n1
U
2
d
x
'H. ).
Contribution of

1
i,e,(
1
:
h
C
0
R
n1
1
i,e,(
1
U(kz) = d
N x
h
2
}
0
1
k
n1
1 2}
02
o
(] 1)k
2
= d
N x
h
2
}
0
1
(] 1)k
nC1
(] 1)h
;1
2oA
2
h
2c
|
.
To annihilate this projection, we should nd }
1
such that
R
n1
1
i,e,(
1
U(kz)
R
n1
1
i,e
U(kz) = 0.
This is equivalent to
T}
1
:= d
N x
h
2
}
0
1
(] 1)k
nC1
(] 1)h
;1
2oA
2
h
2c
|
= 2A
] 1
20
1
d
N x
'H. ).
(3.24)
Hence it is sufcient to set
}
0
1
=
2A(] 1)k
nC1
(] 1)h
;C1
2oA
2
h
2cC2
] 1
20
1
'H. )
c
(] 1)k
nC1
(] 1)h
;C1
2oA
2
h
2cC2
.
where c is a constant to be chosen so that

1
0
}
0
1
Js = 0. By (2.16), we have
(] 1)k
nC1
(] 1)h
;C1
2oA
2
h
2cC2
= oh
c1
dh
dA
h
c
.
and so the required condition becomes
c =
] 1
20
1
2A(] 1)
1
0
k
nC1
(;1)h
pC1
2cA
2
h
2C2
'H. )Js
1
0
oh
c1
Jh
JA
h
c
Js
.
As we can easily check from (2.14) and (2.19), c coincides with A
0
and therefore
we have in conclusion
(3.25)
}
0
1
=
2A(] 1)k
nC1
(] 1)h
;C1
2oA
2
h
2cC2
] 1
20
1
'H. )
A
0
(] 1)k
nC1
(] 1)h
;C1
2oA
2
h
2cC2
.
We evaluate next the projection of the

1s in (3.23) onto the kernel of L
i
. This
corresponds to multiplying the error terms by d
n
U(k), m = 1. . . . . n 1, inte-
grating over R
n1
, and taking the real part. As before, we are left to consider only
two terms:

1
i,o
and

1
i,o,(
1
. The nal result (proven in section 4 of [36]) is the
following:
(3.26)
R
n1
(

1
i,o

1
i,o,(
1
)d
n
U(k(cs)z)Jz
=
] 1
20
1
hk
C
0
h
0
2A
2
0
] 1
h
c
00
n
0
h
01
2A
2
o
] 1
h
c1
h
0
0
n

0
] 1
h
c
((V
1
)
2
V )
n
1
2
h
0
2A
2
0
] 1
h
c
}
(d
2
}n
g
11
)
}
2AA
0
1
(0 o)h
;1
(] 1)h
0
2oA
2
h
c
|
H
n
H
n
'H. )
(] 1)(3
c
0
)h
20
16c0A
4
;1
h
2c
2A
2
(5o 30)h
0Cc
(] 1)h
0
2A
2
oh
c
.
We notice that the operator between brackets coincides precisely with the one in
(2.22), corresponding to the second variation of the reduced functional that we
determined in Subsection 2.5. This is going to be useful in the last section to get
full solvability.
Remark 3.4. According to the considerations in Subsection 2.4, to every normal
variation of , there corresponds some variation in the phase due to both the vari-
ation of position and the variation of the constant A. Recall that the phase of the
approximate solution is the following:
J
t
=
1
c
}(cs) =
1
c
tx
0
}
0
Jl.
Differentiating with respect to a variation v (see (1.12)) we obtain
d
dv
J
t
=
1
c
tx
0
A
h
c
1
c
tx
0
Aoh
c1
dh
dA
A

dh
dV
dV
dv
1
2
1
c
tx
0
d
g
11
Ah
c
.
Recalling formula (2.17), we nd
d
dv
J
t
=
1
c
] 1
2A
A
tx
0
h
01
dh
dA
1
c
tx
0
Aoh
c1
dh
dV
dV
dv
1
2
1
c
tx
0
d
g
11
Ah
c
.
Therefore, when we take a variation v
2
of ,, this also corresponds to a variation
of the phase of
J
J
2
J
t
. Notice that multiplying the horizontal part by hd
n
U corre-
sponds to adding a variation of
2
k
.
Hence, integrating by parts, we get
A
] 1
20
1

n
d
n
g
11
] 1
2A
A
2
h
01
dh
dA
Aoh
c1
dh
dV
dV
dv
2
1
2
Ah
c
d
2
g
11
.
Remark 3.5. If we multiply the operators J and T (see (2.22) and (3.24)) by
h( s)k( s) and h( s), respectively, they become self-adjoint. This fact will become
crucial below; see in particular Subsection 4.3.
4 Lyapunov-Schmidt Reduction of the Problem
In this section we show how to reduce problem (1.16) to a system of three ordi-
nary (integro)differential equations on R0. 1|. We rst introduce a metric on the
normal bundle N,
t
of ,
t
and then study operators that mimic the properties of the
linearization of (1.16) near an approximate solution. Next, we turn to the reduction
procedure: this follows basically from a localization method, since the functions
we are dealing with have an exponential decay away from ,
t
. We introduce a set
1
consisting of approximate (resonant) eigenfunctions of the linearized operator

1
t
: calling

H
t
the orthogonal complement of this set (which has to be multiplied
by a phase factor close to e
i(((tx))t
), we show in Proposition 4.14 that 1
t
is
invertible on the projection onto this set once suitable weighted norms are intro-
duced.
4.1 A Metric Structure on N
"
In this subsection we dene a metric g
t
on N,
t
, the normal bundle to ,
t
, and
then introduce some basic tools that are useful for working in local coordinates on
this set.
First of all, we choose a local orthonormal frame (1
i
)
i
in N, and, using the
notation of [34, subsection 2.2], we set V
1
J
N s
1
}
=
/
}
(d
N x
)1
/
, ;. l = 1. . . . . n1. If
we impose that the 1
}
s are transported in parallel via the normal connection V
1
,
as in Subsection 2.1, we nd that
/
}
(d
N x
) 0 for all ;. l. As a consequence (see
formula (18) in [34]), we have that if (V
}
)
}
, ; = 1. . . . . n 1, is a normal section
to ,, then the components of the normal Laplacian ^
1
V are simply given by
(4.1) (^
1
V )
}
= ^
,
(V
}
) = d
2
N xN x
V
}
. ; = 1. . . . . n 1.
We next dene a metric g on N, as follows. Given N,, a tangent vector
W T
N, can be identied with the velocity of a curve n(t ) in N, that is equal

to at time t = 0. The metric g on N, acts on an arbitrary couple (W.

W)
(T
N,)
2
in the following way (see [20, p. 79])
g(W.

W) = g(
W.

W)
D
1
n
Jt
tD0
.
D
1
n
Jt
tD0
1
.
In this formula denotes the natural projection from N, onto ,, D
1
nJt the
(normal) covariant derivative of the vector eld n(t ) along the curve n(t ), and
n(t ) a curve in N, with initial value and initial velocity equal to

W.
Following the notation in Subsection 2.1 we have that, if n(t ) = n
}
(t )1
}
(t ),
then
D
1
n
Jt
=
Jn
}
(t )
Jt
1
}
(t ).
Therefore, if we choose a system of coordinates ( s. .) on N, dened by
( s. .) R R
n1
.
}
1
}
(,( s)).
we get that
g
11
( s. .) = g
11
( s) .
/
.
}
V
1
T
1
1
/
. V
1
T
1
1
}
1
= g
11
( s) 1
and
g
1
N
/
( s. .) 0. g
N
/
N
}
( s. .) =
N
/
N
}
.
where we have set d
N
/
= dd .
/
. We also notice that the following co-area type
formula holds for any smooth compactly supported function } : N, R:
(4.2)
1,
} JV
y v
=
,

1,(N x)
}( .)J .
J s.
This follows immediately from the fact that det g = det g and by our choice of
( s. .).
Since in the above coordinates the metric g is diagonal, the Laplacian of any
(real- or complex-valued) function dened on N, with respect to this metric is
^
y v
= d
2
N xN x
d
2
N
}
N
}
in N,.
We next endow N,
t
with a natural metric, inherited by g through a scaling. If
T
t
denotes the dilation . c., we dene a metric g
t
on N,
t
simply by
g
t
=
1
c
2
(T
t
)
g|.
In particular, choosing coordinates (s. .) on N,
t
via the scaling ( s. .) = c(s. .),
we easily check that the components of g
t
are given by
( g
t
)
11
(s. .) = g
11
( s) 1. ( g
t
)
1/
(s. .) 0. ( g
t
)
/}
(s. .) =
/}
.
Therefore, if [ is a smooth function in N,
t
, it follows that in the above coordinates
(s. .)
^
y v
"
[ = d
2
xx
[ d
2
}}
[ in N,
t
.
In the case [(s. .) = e
i
y
( x
u(s. .), for

} = A
h
c
(see (1.9)) and for u real, we
clearly have that
^
y v
"
[ = e
i
y
( x
d
2
xx
u 2i

} e
i
y
( x
d
x
u

}
2
e
i
y
( x
u e
i
y
( x
d
2
}}
u.
Similarly to (4.2), we easily nd that
(4.3)
1,
"
} JV
y v
"
=
,
"

1,
"
(x)
}(.)J.
Js.
4.2 Localizing the Problem to a Subset of the Normal Bundle N
"
We next exploit the exponential decay of solutions (or approximate solutions)
away from ,
t
to reduce (1.16) from the whole scaled manifold M
t
to the normal
bundle N,
t
. This step of the proof closely follows a procedure in [19]. We rst
dene a smooth, nonincreasing cutoff function n : R R satisfying
n(t ) = 1 for t _ 0.
n(t ) = 0 for t _ 1.
n(t ) 0. 1| for every t R.
Next, if (s. .) are the coordinates introduced above in N,
t
, and if (cs) is a section
of N,, using the notation of Subsection 3.1, we dene
z = . (cs).
We will assume throughout the paper that satises the following bounds:
(4.4) ||
1
|
0
|
1
c|
00
|
1
_ Cc
for some xed constant C > 0. Next, for a small

> 0 and for a smooth function
1(cs) > 0, both to be determined below, and for h. k : 0. 1| R as dened in
the introduction, we set
[
0,t
= n
t
(s. z)[
0,t
:= n
1(cs)
[z[
c
1(cs)
e
i
Q
f ."s/
"
h(cs)U(k(cs)z).
(4.5)
where

} (to be dened later) is close to the function } (also dened in the intro-
duction). For t (0. 1), we let S
t
: C
2,r
(M
t
) C
r
(M
t
) be the operator
(4.6) S
t
([) = ^
v
"
[ V(c.)[ [[[
;1
[ in M
t
.
If we let

[
t
denote an approximate solution of (1.16) (we will later take

[
t
equal
to

[
0,t
with some small correction), then setting [ =

[
t

, we have S
t
([) = 0
if and only if
1
t
(
) = S
t
(

[
t
) N
t
(
) in M
t
.
where 1
t
(
) stands for the linear correction in

, namely,
1
t
(
) = ^
v
"
V(c.)
[

[
t
[
;1
(] 1)[

[
t
[
;3
[
t
m(

[
t
) in M
t
.
(4.7)
and where the nonlinear operator N
t
(
) is dened as
N
t
(
) = [

[
t

[
;1
(

[
t

) [

[
t
[
;1
[
t
[

[
t
[
;1
(] 1)[

[
t
[
;3
[
t
m(

[
t
).
(4.8)
Then, in the coordinates (s. z), we can write
= n
t
(z) c
where, with an abuse of notation, we assume dened on N,
t
(through the expo-
nential map normal to ,
t
) and where the correction c is dened on the whole M
t
.
In this way we need to solve the equation
(4.9) 1
t
( n
t
(z)) 1
t
(c) = S
t
(

[
t
) N
t
( n
t
(z) c) in M
t
.
We will require to be supported in a cylindrically shaped region in N,
t
cen-
tered around the zero section. For technical reasons, convenient for proving the
results in the next subsection, we dene
(4.10)

D
t
=
(s. z) N,
t
: [z[ _
c
1
1(cs)
.
and then the subspace of functions in N,
t
as
H
Q
T
"
= {u 1
2
(N,
t
: C) : u is supported in

D
t
].
Using elementary computations, we see that (4.9) is satised if (tautologically) the
following two conditions are imposed:
1
t
() =

S
t
(

[
t
) N
t
( n
t
(z) c)
[

[
t
[
;1
c
(] 1)[

[
t
[
;3
[
t
m(

[
t
c) in

D
t
. H
Q
T
"
.
(4.11)
L
Q
v
"
c = (1 n
t
(z))
S
t
(

[
t
) N
t
( n
t
(z) c)
2V
v
"
n
t
(z) V
v
"
^
v
"
n
t
(z) in M
t
.
(4.12)
where
L
Q
v
"
c = ^
v
"
c V(c.)c
(1 n
t
(z))
[

[
t
[
;1
c (] 1)[

[
t
[
;2
[
t
m(

[
t
c)
.
(4.13)
We next have an existence result for equation (4.12). In order to state it, we
need to introduce some notation. For a regular periodic function p : 0. 1| R,
for m N and t (0. 1), we dene the weighted norms
(4.14) |c|
C
m;
p
= sup
x2
Q
T
"
e
p(tx)j:j
|c|
C
m;
(B
1
(x))
|. . = (s. z).
We also recall the denition of k(cs) in (1.11).
PROPOSITION 4.1 Let k
2
( s) < k
1
( s) < k
0
( s). 1( s) be smooth positive 1-peri-
odic functions in s, and t (0. 1). Then, if V( s). 1
2
( s) > k
2
2
( s) and if
|

[
t
|
C
k
0
. |S
t
(

[
t
)|
C
k
0
_ 1, there exists a positive constant C depending on

,
t, k, k
0
, k
1
, and k
2
such that given any with ||
C
1;
k
1
_ 1, problem (4.12) has a
unique solution c() whose restriction to

D
t
satises
|c()|
C
k
2
_ C
e
t
inf
k
2
Ck
0
K
|S
t
(

[
t
)|
C
k
0
e
t
inf
k
1
Ck
2
K
||
C
1;
k
1
.
(4.15)
Moreover, if

[
1
t
.

[
2
t
satisfy |S
t
(

[
}
t
)|
C
k
0
_ 1, ; = 1. 2, if |
}
|
C
1;
k
1
_ 1, ; =
1. 2. and if c
}
(
}
), ; = 1. 2, are the corresponding solutions, for the restrictions
to

D
t
, we also have
|c(
1
) c(
2
)|
C
k
2
_ C
e
t
inf
k
2
Ck
0
K
S
t
(

[
1
t
) S
t
(

[
2
t
)
k
0
e
t
inf
k
1
Ck
2
K
|
1

2
|
C
1;
k
1
.
(4.16)
Remark 4.2.
(1) The choice of the norm in (4.14) is done for considering functions that
grow at most like e
p(tx)j:j
, and in particular functions that decay at innity if p is
positive. In the left-hand side of (4.15) we have a negative exponent, representing
the fact that c can grow as [z[ increases. However (we will later take k
0
. k
1
. k
2
,
and 1 very close), the coefcients in the right-hand side are so tiny that c is every-
where small in

D
t
, and indeed with an even smaller bound for [z[ close to 0. This
reects the fact that the support of the right-hand side in (4.12) is
c
1(cs)
< [z[ <
c
1
1(cs)
.
so c should decay away from this set.
(2) We introduced the functions k
0
. k
1
, and k
2
for technical reasons, since we
want to allow some exibility for the (exponential) decay rate in [z[.
PROOF: We prove the result only when the manifold M in (NLS
t
) is compact.
For the modications needed for M = R
n
, see Remark 4.3(2).
Consider a smooth, nondecreasing cutoff function , : 0. 1| 0. 1| satisfying
,(t ) = 0 for t _
1
4
.
,(t ) = t for t _
3
4
.
0 _ ,
0
(t ) _ 4 for all t.
0 _ ,
00
(t ) _ 16 for all t.
Next, given a large constant B (to be specied later) depending only on V and k
2
,
we dene ,( s. [z[) as
,( s. [z[) =
B,
j:j
B
for [z[ _ B.
[z[ for B _ [z[ _
t
k
2
(N x)
1.
[z[
t
k
2
(N x)

1
2
2,
[z[
t
k
2
(N x)

1
2
for
t
k
2
(N x)
1 _ [z[ _
t
k
2
(N x)
1.
2
t
k
2
(N x)
[z[ for
t
k
2
(N x)
1 _ [z[ _ 2
t
k
2
(N x)
B.
B,
2t
k
2
(N x)j:j
B
for 2
t
k
2
(N x)
B _ [z[ _ 2
t
k
2
(N x)
.
0 for [z[ _ 2
t
k
2
(N x)
.
By our choice of ,, the function , satises the following inequalities (where, here,
the gradient and the Laplacian are taken with respect to the Euclidean metric)
[V
:
,[ _ 1. ^
:
, _
16 4(n 2)
B
.
Using the above coordinates (s. z), we next dene the barrier function u : M
t

R as
u(s. z) = e
k
2
(tx) Q ;(tx,:)
for [z[ _ 2
c
k
2
(cs)
.
and we extend u identically equal to 1 elsewhere. By our choice of ,, this function
is indeed smooth and strictly positive on the whole M
t
. We consider next the linear
equation (motivated by (4.13))
L
Q
v
"
c = 0 on M
t
.
where 0 : M
t
R is Hlder-continuous (with supp(0) b

D
t
; see (4.18) below).
Since the operator L
Q
v
"
is uniformly elliptic, the latter equation is (uniquely) solv-
able, and we would next like to derive some pointwise estimates on its solutions.
To this aim we dene
(.) =
c(.)
u(.)
. . M
t
.
With this notation, we have that
uL
Q
v
"
^
v
"
u 2V
v
"
V
v
"
u = 0 on M
t
.
Using the expression of the metric coefcients in the coordinates (s. z) (see Lem-
ma 2.1), (4.4), and the properties of the cutoff function ,, we easily check that
^
v
"
u
_ (k
2
( s)
2
o
B
(1) o
t
(1))u for [z[ _
t
k
2
(N x)
.
= 0 elsewhere.
where o
t
(1) 0 as c 0 and o
B
(1) 0 as B o. Therefore we obtain
that the function satises
L
Q
v
"
k
2
2
( s) o
B
(1) o
t
(1)
_
jj
u
for [z[ _ 2
t
k
2
(N x)
.
L
Q
v
"
=

u
elsewhere.
Since we assumed V( s) > k
2
( s)
2
, we obtain that V k
2
( s)
2
o
B
(1) o
t
(1)
is strictly positive (provided B is sufciently large and c sufciently small) for
[z[ _ 2(c
k
2
( s)), and hence the function satises a uniformly elliptic equation
with a nonnegative coefcient in the zeroth-order term with right-hand side given
by

u
. Therefore from the maximum principle we derive the estimate
max
"
[[ _ C max
"
[0[
u
.
where C depends on the uniform lower bound of the above coefcient. The latter
formula clearly implies
[c(.)[ _ Cu(.) max
"
0
u
for every . M
t
.
We next dene the weighted norm
|c|
n,r,u
:= sup
x2
"
c
u
C
m;
(B
1
(x))
.
which is equivalent (with constants depending on B only) to ||
C
m;
k
2
on the set

D
t
.
Using the explicit form of the function u and standard elliptic regularity results, we
can improve the latter inequality to
(4.17) |c|
2,r,u
_ C|0|
0,r,u
.
The proof of the proposition now follows from this linear estimate and the contrac-
tion mapping theorem; in fact, dening
G
,t
(c) = (1 n
t
(s. z))S
t
(

[
t
) N
t
( n
t
(z) c)|
2V
v
"
n
t
(z) V
v
"
^
v
"
n
t
(z).
equation (4.12) is equivalent to
(4.18) c = L
1
Q
v
"
G
,t
(c).
First of all, notice that L
Q
v
"
is invertible since we are assuming M (and hence
M
t
) to be compact (see the beginning of the proof). Second, to apply (4.17), we
need to estimate |G
,t
(c)|
0,r,u
, together with its Lipschitz dependence in c.
Let us consider, for instance, the term (1 n
t
(s. z))S
t
(

[
t
). Using the fact that
(1 n
t
) is 0 for [z[ _ c
1(cs), that S
t
(

[
t
) is 0 for [z[ _ (c
1)1(cs),
and that k
2
< 1, we obtain
(4.19)
|(1 n
t
)S
t
(

[
t
)|
0,r,u
_ C sup
"
K."s/
j:j
"
C1
K."s/
S
t
(

[
t
)
u
(B
1
(:))
_ Ce
t
inf
k
0
Ck
2
K
|S
t
(

[
t
)|
C
k
0
.
Now to estimate the remaining terms of G
,t
, we notice that
[N
t
( n
t
(z) c)[ _
C[

[
t
[
;2
[ n
t
(z) c[
2
if [ n
t
(z) c[ _ [

[
t
[.
[ n
t
(z) c[
;
otherwise.
Since ] > 1, we can nd a number (0. 1) such that ] 2 1 > 0, so the
last formula implies
(4.20)
[N
t
( n
t
(z) c)[ _ C
[

[
t
[
;1c
([ n
t
(z)[
c
[c[
c
)([ n
t
(z)[ [c[)
[ n
t
(z)[
;
[c[
;
.
Using the fact that |

[
t
|
C
k
0
. ||
C
1;
k
1
_ 1 and reasoning as for (4.19), after some
computations we deduce (assuming |c|
1
_ 1, which will be veried later)
(4.21)
|G
,t
(c)|
0,r,u
_ C
e
t
inf
k
2
Ck
0
K
|S
t
(

[
t
)|
C
k
0
e
t
inf
k
2
Ck
1
K
||
C
1;
k
1
e
t
inf
pk
1
Ck
2
K
||
C
0;
k
1
e
t
inf
.p1/k
0
Ck
1
Ck
2
K
||
C
0;
k
1
e
t
(;1c) inf
k
0
K
|c|
;1
1
|c|
0,r,u
.
Similarly, for two functions c
1
and c
2
with |c
1
|
1
. |c
2
|
1
_ 1 and with nite
| |
0,r,u
norm, we have
(4.22) |G
,t
(c
1
) G
,t
(c
2
)|
0,r,u
_
e
t
(;1c) inf
k
0
K
|c
1
|
;1
1
|c
2
|
;1
1
|c
1
c
2
|
0,r,u
.
We now consider the map c G
,t
(c) in the set
B =
c : |c|
0,r,u
_
C
1
e
t
inf
k
2
Ck
0
K
|S
t
(

[
t
)|
C
k
0
e
t
inf
k
2
Ck
1
K
||
C
1;
k
1
.
where C
1
is a sufciently large positive constant. Notice that if c B, then
|c|
1
= o
t
(1). From (4.21) and (4.22), it then follows that this map is a contrac-
tion from B into itself, endowed with the above norm, and therefore a solution c
exists as a xed point of G
,t
. The fact that k
2
< 1 implies that the norm| |
C
k
2
is equivalent to | |
0,r,u
in

D
t
(see also the comments in Remark 4.2), so we ob-
tain (4.15). A similar reasoning, still based on regularity theory and elementary
inequalities, also yields (4.16).
Remark 4.3.
(1) From elliptic regularity theory it follows that in (4.15)(4.16) the norm
| |
C
k
2
can be replaced by the stronger | |
C
2;
k
2
, yielding an estimate in the norm
| |
C
0
k
2
for any t
0
(t. 1).
(2) In the case M = R
n
, the above proof needs to be slightly modied: in
fact, the invertibility of L
Q
v
"
will be guaranteed provided we work in an appro-
priate class of functions Y decaying exponentially at innity. To guarantee this
condition, we can vary the form of the barrier function u in order that it both re-
mains a supersolution of L
Q
v
"
= 0 and decays exponentially to 0 at innity. This is
indeed possible using the uniform positive lower bound on V ; see (1.1). We omit
the details of this construction.
As a consequence of Proposition 4.1, we obtain that the solvability of (NLS
t
) is
equivalent to that of (4.11).
PROPOSITION 4.4 Suppose the assumptions of Proposition 4.1 hold, and consider
the corresponding c = c(). Then [ =

[
t
n
t
(z) c() solves (1.16) if and
only if H
Q
T
"
satises
(4.23) 1
t
() =

S
t
() in

D
t
.
with
S
t
() = S
t
(

[
t
) N
t
( n
t
(z) c()) [

[
t
[
;1
c() (4.24)
(] 1)[

[
t
[
;3
[
t
m(

[
t

c()) in

D
t
.
where n
t
. S
t
, and N
t
are given in (4.5), (4.6), and (4.8), respectively.
4.3 Construction of an Approximate Kernel for L
"
We perform here some preliminary analysis useful to understand the spectral
properties of 1
t
. More precisely, we consider a model case, when the domain

D
t
(see (4.10)) is replaced by 0.
1
t
| R
n1
and the prole of approximate solutions
is independent of the variable s (only the phase varies, periodically in s). As in
formula (1.9), we consider positive constants

V .

h, and

k satisfying
(4.25)

h = (

}
2

V )
1
p1
.

k = (

}
2

V )
1
2
.
Our goal is to study the following eigenvalue problem, which models our linearized
equation:
1
t
u = zu in 0.
1
t
| R
n1
.
1
t
u = ^
y v
"
u

V u

h
;1
U(
k.)
;1
u
(] 1)
h
;1
U(
k.)
;1
e
i
y
( x
m(e
i
y
( x
u).
(4.26)
and in particular we would like to characterize the small eigenvalues and the cor-
responding eigenfunctions.
First of all, we can write u as
u = e
i
y
( x
(u
i
i u
i
)
for some real u
i
and u
i
. With this notation, we are reduced to studying the coupled
system
^
y v
"
u
i
(
V

}
2
)u
i
]
h
;1
U(
k.)
;1
u
i
2

}
Ju
i
Jx
= zu
i
in 0.
1
t
| R
n1
.
^
y v
"
u
i
(
V

}
2
)u
i

h
;1
U(
k.)
;1
u
i
2

}
Ju
r
Jx
= zu
i
in 0.
1
t
| R
n1
.
Making the change of variables .

k. and using (4.25), we are reduced to
(4.27)
1
y
k
2
J
2
u
r
Jx
2
^
+
u
i
u
i
]U(.)
;1
u
i

2
y
(
y
k
2
Ju
i
Jx
=
2
y
k
2
u
i
in 0.
1
t
| R
n1
.
1
y
k
2
J
2
u
i
Jx
2
^
+
u
i
u
i
U(.)
;1
u
i

2
y
(
y
k
2
Ju
r
Jx
=
2
y
k
2
u
i
in 0.
1
t
| R
n1
.
It is now convenient to use a Fourier decomposition in s of u
i
and u
i
, writing
u
i
=
cos
2c;s
1
u
i,c,}
(.) sin
2c;s
1
u
i,x,}
(.)
.
u
i
=
cos
2c;s
1
u
i,c,}
(.) sin
2c;s
1
u
i,x,}
(.)
.
where s 0.
1
t
| and . R
n1
. In this way the functions u
i,c,}
. u
i,x,}
. u
i,c,}
, and
u
i,x,}
satisfy the following systems of equations:
^
+
u
i,c,}

1
4t
2
t
2
}
2
1
2 y
k
2
u
i,c,}
]U(.)
;1
u
i,c,}

4t
y
( t}
1
y
k
2
u
i,x,}
=
2
y
k
2
u
i,c,}
in R
n1
.
^
+
u
i,x,}

1
4t
2
t
2
}
2
1
2 y
k
2
u
i,x,}
U(.)
;1
u
i,x,}

4t
y
( t}
1
y
k
2
u
i,c,}
=
2
y
k
2
u
i,x,}
in R
n1
.
^
+
u
i,x,}

1
4t
2
t
2
}
2
1
2 y
k
2
u
i,x,}
]U(.)
;1
u
i,x,}

4t
y
( t}
1
y
k
2
u
i,c,}
=
2
y
k
2
u
i,x,}
in R
n1
.
^
+
u
i,c,}

1
4t
2
t
2
}
2
1
2 y
k
2
u
i,c,}
U(.)
;1
u
i,c,}

4t
y
( t}
1
y
k
2
u
i,x,}
=
2
y
k
2
u
i,c,}
in R
n1
.
If we set
2c;
1
k
= .
2

}
k
= u. and

z =
z
k
2
.
then the latter two systems are equivalent to the following:
(4.28)
^
+
u (1
2
)u ]U(.)
;1
u u =

zu in R
n1
.
^
+
(1
2
) U(.)
;1
uu =

z in R
n1
.
The equivalence with the second system is obvious: for the rst one it is suf-
cient to switch the sign of the second component. We characterize the spectrum of
the last system in the next proposition: the value of u is xed, while is allowed to
vary. We remark that it is irrelevant for our purposes to take positive or negative,
since we can still switch the sign of one of the two components.
PROPOSITION 4.5 Let n
. o
, and t
denote the rst three eigenvalues of (4.28).

Then there exists u
0
> 0 such that for u 0. u
0
| the following properties hold:
(i) There exists
0
> 0 such that n
is simple, increasing, and differentiable

in for 0.
0
|,
Jn
J
> 0 for (0.
0
|, n
0
< 0, and n
0
> 0.
(ii) The eigenvalue o
is 0 for = 0 with multiplicity n, it satises

Jc
J
> 0
for small and positive, and stays uniformly bounded away from 0 if
stays bounded away from 0.
(iii) t
is strictly positive and stays uniformly bounded away from 0 for all s.
(iv) The eigenfunction u
corresponding to n
is simple, radial in ., and radi-

ally decreasing and depends smoothly on : for = 0 the eigenfunction
of (4.28) corresponding to n
0
< 0 is of the form (

7. 0) with

7 radial
and radially decreasing, while those corresponding to o
0
= 0 are linear
combinations of (V
+
j
U. 0), ; = 1. . . . . n 1, and (0. U).
(v) Let be the unique for which n
N
= 0 (see (i)); then the corresponding
eigenfunction is of the form (7. W) for some radial functions 7 and W
satisfying the decay [7[[W[ _ Ce
(1Cy n)j+j
for some constants C. n > 0.
PROOF: This result is known for u = 0; see, for example, [34, prop. 4.2] and
[39, prop. 2.9].
For u = 0 sufciently small, the functions n
, o
, and t
will be C
1
-close to those corresponding to u = 0; therefore, to prove (i)(iv), it
is sufcient to show that n
and o
are twice-differentiable in for small, that

dn
d = do
d = 0, and that d
2
n
d
2
. d
2
o
d
2
> 0.
We prove this statement only heuristically, but a rigorous proof can easily be
derived. Differentiating
(4.29)
^
+
u
(1
2
)u
]U(.)
;1
u
= n
in R
n1
.
^
+
(1
2
)
U(.)
;1
uu
= n
in R
n1
.
with respect to , we nd
(4.30)
^
+
Ju
J
(1
2
)
Ju
J
]U(.)
;1 Ju
J
u
J
J
2u
= n
Ju
J

Jn
J
u
in R
n1
.
^
+
J
J
(1
2
)
J
J
U(.)
;1 J
J
u
Ju
J
2
uu
= n
J

Jn
in R
n1
.
To compute
Jn
J
at = 0 it is sufcient to multiply the rst equation by u
and
the second by
to take the sum and integrate; if we choose

Ju
J
and
J
J
so that
R
n1
u
Ju
J
= 0 (choosing, e.g.,

(u
2
) = 1 for all s), then with

an integration by parts we nd that
dn
D0
R
n1
u
2
= 2u
R
n1
u
0
0
.
Using the fact that
0
= 0 (see (iv)), we then obtain
Jn
J
[
D0
= 0. The same
argument applies for evaluating
Jc
J
[
D0
, since the eigenfunctions corresponding
to o
0
= 0 always have one component vanishing.
To compute the second derivative with respect to , we differentiate (4.30) once
more at = 0, obtaining
^
+
J
2
u
J
2

J
2
u
J
2
]U(.)
;1 J
2
u
J
2
2u
J
J
2u
0
=
J
2
n
J
2
u
0
in R
n1
.
^
+
J
2
J
2

J
2
J
2
U(.)
;1 J
2
J
2
2u
Ju
J
2
0
=
J
2
n
J
2

0
in R
n1
.
As for the previous case we get
d
2
n
d
2
D0
= 2 2u
R
n1
u
0
d
D0
0
du
D0
:
so, using the smallness of u, the claim follows.
For the second derivative of o
the procedure is similar, but notice that in this

case we might obtain a multivalued function, due to the multiplicity (n) of o
0
(see (ii)). However, if in the last formula we plug in the corresponding eigen-
functions (see (iv)), we still obtain a sign condition for each of the two branches
of o
(one of them will have multiplicity n 1 by the rotation invariance of the

equations).
Remark 4.6. Using the same argument in the previous proof, we can show that
dn
d
DN
= 2 2u
R
n1
7
N
W
N
.
Remark 4.7. Proposition 4.5 is the only result where the smallness of the constant
A is used; see Theorem 1.1. We remark that V

V implies

} = A
h
c
, and
that u = 2(

}
k), so the smallness of A is equivalent to that of u. Notice that by

(1.1) and (1.9), when A 0,

h and

k stay uniformly bounded and bounded away
from 0.
We believe that dropping this smallness condition might lead to further reso-
nance phenomena in addition to those encountered here (see the introduction and
the last section).
Remark 4.8. Considering (4.30) with o
replacing n
and for = 0, we nd that

L
0
i
Ju
J
[
D0
= u
0
and L
0
i
J
J
[
D0
= uu
0
, where
L
0
i
= ^
+
]U(.)
;1
.
L
0
i
= ^
+
U(.)
;1
.
Since for = 0 we have (u
0
.
0
) = (d
}
U. 0) or (u
0
.
0
) = (0. U) (see (iv)) and
(4.31) L
0
i
1
] 1
U
1
2
VU(.) .
= U. L
0
i
(.
}
U(.)) = 2d
}
U.
see Subsection 3.2, we nd that
d
D0
=
u
2
.
}
U(.).
du
D0
= u
1
] 1
U
1
2
VU(.) .
.
These expressions, together with (3.26) and some integration by parts, allow us
to compute explicitly d
2
o
d
2
, whose values along the two branches are
d
2
o
d
2
=
2
(] 1)
(] 1) 2A
2
0
h
2c;C1
.
d
2
o
d
2
=
2
(] 1)
(] 1) 2A
2
o
h
2c;C1
.
Therefore, we nd that the second derivatives of the eigenfunctions satisfy, respec-
tively, the equations
L
0
i
d
2
u
d
2
=
2
] 1
(] 1) 2A
2
0
h
2c;C1
V
}
U
2V
}
U 4A
2
h
2c;C1
.
}
U.
(4.32)
L
0
i
d
2
d
2
=
2
] 1
(] 1) 2A
2
o
h
2c;C1
U
2U 8A
2
h
2c;C1
U.
(4.33)
These formulas will be crucial later on. Below, we will denote for brevity
(4.34)

V
}
:=
1
2
d
2
u
d
2
. ; = 1. . . . . n 1.

W :=
1
2
d
2
d
2
.
The factor
1
2
arises in the Taylor expansion of the eigenfunctions in , and ; is the
index in (4.32).
We next consider the case of variable coefcients, which can be reduced to the
previous one through a localization argument in s. To have a more accurate model
for 1
t
, the constants

k and

} in (4.27) have to be substituted with the functions
k(cs) and }(cs) satisfying (1.11). Specically, in N,
t
we dene
(4.35)
1
1
t
u = ^
y v
"
u V(cs)u h(cs)
;1
U(k(cs).)
;1
u
(] 1)h(cs)
;1
U(k(cs).)
;1
e
i
f."s/
"
m(e
i
f."s/
"
u)
(recall the denition of g
t
in Subsection 4.1; in particular, working with the coor-
dinates (s. .) integrals will be computed using the co-area formula (4.3)).
Before proving rigorous results, we rst discuss heuristically what the approx-
imate kernel of 1
1
t
should look like. Using Fourier expansions as above (freezing
the coefcients at some s), the prole of the functions that lie in an approximate
kernel of 1
1
t
will be given by the solution of (recall (4.28))
(4.36)
^
+
u (1
2
)u ]U(.)
;1
u 2
(
0
(N x)
k(N x)
=

zu in R
n1
.
^
+
(1
2
) U(.)
;1
2
(
0
(N x)
k(N x)
u =

z in R
n1
.
where

z is close to 0. For small (low Fourier modes), Proposition 4.5(iv) gives
the prole V
+
U(k( s).) or i U(k( s).) (recall also the scaling in . before (4.27)).
The remaining part of the approximate kernel is the counterpart of that given in
Proposition 4.5(v); for variable coefcients it is a uniquely dened function ( s)
such that
(4.37) n
(N x)
= 0.
where n
here stands for the rst eigenvalue of (4.36). We denote by
7
(N x)
(k( s).). W
(N x)
(k( s).)
the components of the relative eigenfunction.

We next consider two bases of eigenfunctions for the weighted eigenvalue prob-
lems (the operators J and T are dened in (2.22) and (6.18) and are self-adjoint)
(4.38) Jc
}
( s) = h( s)
0
z
}
c
}
( s). To
}
= h( s)
c
j
}
o
}
.
Because of the weights on the right-hand sides, we can choose these eigenfunctions
to be normalized so that
1
0
h
0
c
}
c
/
=
}/
and
1
0
h
c
o
}
o
/
=
}/
; this choice will
be useful in Subsection 6.2.
These heuristic arguments suggest that the following subspaces 1
1,
and 1
2,
,
where is a small positive constant, once multiplied by e
i((tx)t
, consist of ap-
proximate eigenfunctions for 1
1
t
with eigenvalues close to 0 (this will be veried
below, in the proof of Proposition 4.9; see also Remark 4.11):
1
1,
= span
h(cs)
pC1
4
'c
}
(cs). V
+
U(k.)) i c'c
0
}
(cs). .)
}
0
k
U(k.)
c
2
k
2
'c
00
}
(cs). V(k.))
.
(4.39)
1
2,
= span
h(cs)
1
2
o
}
(cs)i U(k.)
2c}
0
(cs)
k
o
0
}
(cs)

U(.)
c
2
k
2
o
00
}
(cs)W(k.)
.
(4.40)
; = 0. . . . .

t
, where
U =
1
h
;1
(] 1)
U(k.)
1
2k
VU(k.) .
.
Here V = (V
}
)
}D1,...,n1
is the counterpart of

V in (4.34) substituting

h with
h( s) (the same holds for W). The choice of the weights (as powers of h) in (4.38)
and in 1
1,
and 1
2,
are again done for technical reasons, and will be convenient
below; see, in particular, Subsection 6.2.
We also need to construct an approximate kernel with the prole (7. W); see
Proposition 4.5(v). To this aim we introduce the functions (recall (4.37))
(4.41)
O
1,
( s) =
R
n1
7
2
(N x)
. O
2,
( s) =
R
n1
W
2
(N x)
.
O
3,
( s) =
R
n1
7
(N x)
W
(N x)
.
and consider the following eigenvalue problem (with periodic boundary condi-
tions)
c
2
00
k
2
2
=
v
1 2}
0
O
3;
k
on 0. 1|.
By Weyls asymptotic formula we have that the eigenvalues v
}
(counted with mul-
tiplicity) have the qualitative behavior v
}
. 1c
2
;
2
as ; o. Hence, there
is a rst index ;
t
(of order
1
t
) for which v
}
"
_ 0. Setting v
}
= v
}
"
C}
and denoting
by
}
the eigenfunctions corresponding to v
}
, we then have
(4.42)
c
2
00
}
k
2
}
=
v
}
1 2}
0
O
3;
k
}
.
v
}
=

C
0
c; O(c
2
;
2
) O(c) if [; [ _

2
c
.
where > 0 is any given positive (small) constant. Notice that the family (
}
)
}
can be chosen normalized in 1
2
with the weight
1
1 2}
0
O
3;
k
(this follows from (4.42) and the Courant-Fischer formula). Next we set
(4.43)
}
=
1
k
1
O
1,
k
2
2
2}
0
kO
3,
v
}
c
0
}
.
By our choices, the functions
}
and
}
satisfy (this system will be useful in Sub-
section 6.3) for [; [ _
2
c
(4.44)
c
2
00
}
k
2
}
2}
0
O
3;
O
1;
(c
0
}
k
}
) =
v
}
}
(O(v
2
}
) O(c))
}
.
c
2
00
}
k
2
}
2}
0
O
3;
O
2;
(c
0
}
k
}
) =
v
}
}
(O(v
2
}
) O(c))
}
.
Our next goal is to introduce a family of approximate eigenfunctions of 1
1
t
that
have the prole (from now on, we might denote (7
(N x)
. W
(N x)
), see (4.37), simply
by (7
. W
))
(4.45)
3,}
:= (
}
q
}
)7
,
}
d7
d
i
}
W
i k
}
dW
d
:
the functions
}
and
}
are as in (4.42)(4.43), while q
}
, ,
}
, and k
}
are small cor-
rections to be chosen properly so that 1
1
t
(e
i((tx)t]
3,}
) = v
}
e
i((tx)t]
3,}
,
up to an error o(v
}
) O(c).
With simple computations, using (4.29), (4.30), Remark 4.6, and (4.44), we nd
that
e
i
f."s/
"
1
1
t
(e
i
f."s/
"
3,}
) v
}
e
i
f."s/
"
3,}
= 7
2}
0
O
3,
O
1,
2O
1,
,
}
k (c
0
}
k
}
)|
W
2}
0
(,
}
k (c
0
}

}
k) kq
}
)
d7
d
(c
2
,
00
}

2
k
2
,
}
)
dW
d
2}
0
(ck
0
}
k,
}
)
i7
2}
0
(k
}
k k
}
c
0
}
cq
0
}
)
iW
2}
0
O
3,
O
2,
(2O
2,
kk
}
k
}
c
0
}
)
i
d7
d
2}
0
(kk
}
c,
0
}
) i
dW
d
(c
2
k
00
}

2
k
2
k
}
)
7
(c
2
q
00
}

2
k
2
q
}
) (O(v
2
}
) O(c))
}
(O(v
2
}
) O(c))
}
.
To make the coefcients of the terms 7
, W
, and iW
in the second and fourth

lines vanish, we choose
,
}
=
1
2kO
1,
(c
0
}
k
}
). k
}
=
1
2kO
2,
(k
}
c
0
}
).
q
}
=
1
k
(c
0
}

}
k k,
}
).
Using (4.43) we get
,
}
= c
0
}
1
2k
1
k
2
2
2}
0
kO
3,
v
}
.
k
}
=
1
2k
v
}
1 2}
0
O
3;
k
}
O(v
2
}
)
}
.
(4.46)
These equations and (4.42) imply the relations c
2
,
00
}

2
k
2
,
}
= O(v
2
}
)
}
,
ck
0
}
k,
}
= O(v
2
}
)
}
, kk
}
c,
0
}
= O(v
2
}
)
}
, and c
2
k
00
}

2
k
2
k
}
=
O(v
2
}
)
}
. Similarly, we nd
(4.47) q
}
=
}
1
2k
1
k 2}
0
O
3,
v
}
.
This also yields ,
}
k (c
0
}

}
k) kq
}
= O(v
2
}
)
}
, k
}
k k
}
c
0
}

cq
0
}
= O(v
2
}
)
}
and c
2
q
00
}

2
k
2
q
}
= O(v
2
}
)
}
, so we obtain
1
1
t
(e
i
f."s/
"
3,}
) = v
}
e
i
f."s/
"
3,}
(O(v
2
}
)
O(c))
}
(O(v
2
}
) O(c))
}
for [; [ _

2
c
.
(4.48)
which was our claim. We next dene
1
3,
= span
(
}
q
}
)7
,
}
d7
d
i
}
W
}
i k
}
dW
d
:
; =
2
c
. . . . .
2
c
.
(4.49)
In the 1
/,
s we added some corrections to the approximate eigenfunctions that
take into account the variation of the prole with the frequency; see the derivation
of (4.28) and Remark 4.8. In 1
1,
and 1
2,
the corrections are up to the second
order (in c; ), while in 1
3,
only up to the rst; the reason is that the corresponding
eigenvalues have a quadratic dependence in c; for 1
1,
and 1
2,
(they correspond
to n
in Proposition 4.5), and an afne dependence in c; for 1

3,
(corresponding
to u
in Proposition 4.5). Since the former dependence is more delicate in the

indices, we need a more accurate expansion of the eigenfunctions. We nally set
(4.50) 1
= span{1
1,
. 1
2,
. 1
3,
].
4.4 Invertibility of L
"
in the Orthogonal Complement of K
Since 1
(multiplied by e
i(((tx))t
) is a good candidate for the span of the
eigenfunctions of 1
1
t
with small eigenvalues, it seems plausible to invert 1
1
t
on the
orthogonal complement to e
i(((tx))t
1
; this is the content of the next result. We

recall the denition of the constant Ain the introduction.
PROPOSITION 4.9 There exists A
0
sufciently small such that for any A 0. A
0
|
the following property holds: For > 0 small enough there exists C > 0 (indepen-
dent of ) such that if
(4.51) m
1,
"
e
i
f."s/
"

JV
y v
"
= 0 for all 1
.
we have |
t
1
1
t
()|
1
2
(1,
"
)
_ C
1
2
||
1
2
(1,
"
)
. Here
t
denotes the projec-
tion in 1
2
(N,
t
) onto the orthogonal complement of the set {e
i(((tx))t
:
1
].
Before starting the proof, which relies on a localization argument and the spec-
tral analysis of Proposition 4.5, we introduce some notation and a preliminary
lemma. We x s 0. 1| and denote by

} ,

h,

k, and the values of }
0
(s),
h(s), k(s), and (s), respectively, so that the counterpart of (1.11) holds true. For
a large constant

C
0
to be xed later, we also dene
1
1,
= span
' c
}
(cs). V
+
U(
k.))
i c' c
0
}
(cs). .)
k
U(
k.)
c
2
k
2
'c
00
}
(cs).

V(
k.))
1
2,
= span
o
}
(cs)i U(
k.)
2c}
0
(cs)
k
o
0
}
(cs)
U(.)
c
2
k
2
o
00
}
(cs)

W(
k.)
.
; = 0. . . . . (

C
0
c), and
1
3,1,
= span
cos(
}
s)7
y
j
(
k.) i sin(
}
s)W
y
j
(
k.) :
; =

2
C
0
c
. . . . .
C
0
c
1
3,2,
= span
sin(
}
s)7
y
j
(
k.) i cos(
}
s)W
y
j
(
k.) :
; =

2
C
0
c
. . . . .
C
0
c
.
where
U =
h
;1
(] 1)
U(
k.)
1
2
k
VU(
k.) .
.

}
=
k1
2c
;
2c
1
(4.52)
(again, the latter square bracket stands for the integer part, and this choice makes
the functions
1
t
-periodic). In the above formulas ( c
}
)
}
are the eigenfunctions of
the normal Laplacian with the at metric on ,, and o
}
those of d
2
N xN x
on 0. 1|; the
symbols 7
y
j
and W
y
j
stand for the components of the eigenfunctions of (4.28)
corresponding to n
y
j
.
In analogy with (4.50), we also dene
(4.53)

1
= span

1
1,
.

1
2,
.

1
3,1,
.

1
3,2,
.
Given a small constant n > 0 to be chosen later (of order

c), we consider then
a smooth cutoff function ,
n
(depending only on s) with support near
yx
t
and with
length of order
n
t
. For example, we can take ,
n
(s) = ,(
t
n
(s
yx
t
)) for a xed,
compactly supported cutoff , that is 1 in a neighborhood of 0. The next result uses
Fourier cancellation and is related to lemma 2.7 in [39].
LEMMA 4.10 There exists

C
0
sufciently large (depending only on V , 1, and A
0
)
with the following property: for any given integer number m there exists C
n
> 0
depending on m and ,
n
such that for [; [ _
2
(

C
0
c) and for [l[ _
2
c we have
,
n
(s)
/
(s) cos(
}
s)Js
,
n
(s)
/
(s) sin(
}
s)Js
_
1
c
C
n
[v
/
[
n
n(1 [v
/
[)
c
n
n
.
PROOF: We clearly have that (cos(
}
s))
00
=
2
}
cos(
}
s); therefore, integrat-
ing by parts, after some manipulation we obtain that
(4.54)
,
n
/
(s) cos(
}
s)Js
=
1

2
}

k
2

2

l
1C2
y
(
Q
3;y
y
ky
,
n
/
(s) cos(
}
s)
k
2

k
2

2
v
/
1 2

}
O
3;y
y
ky
v
/
1 2}
0
O
3;
k
cos(
}
s),
00
n
/
(s) 2,
0
n
0
/
(s)|
.
By (4.52) the numbers
}
satisfy
}
.

k
2tt}
1
for [; [ _
2
(

C
0
c), while by
(4.42) [v
/
[ _
1
2
C
0
2
for [l[ _
2
c. Notice also that 1 2}
0
O
3;
k
is uniformly
bounded above and below by positive constants when A
0
tends to 0 (see, for ex-
ample, the comments in Remark 4.7). By these facts and the properties of ,
n
we
nd
,
n
/
(s) cos(
}
s)Js
_
1
c
C
[v
/
[
n(1 [v
/
[)
c
n
.
which yields the statement for m = 1 (similar computations can be performed to
deal with the sine function). The factor
1
t
inside the brackets arises from the fact
that we are integrating over the interval 0.
1
t
| and from the normalization of
}
(see
the comments before (4.42)). To obtain the statement for general m, it is sufcient
to iterate the procedure for (4.54) m times and integrate by parts.
PROOF OF PROPOSITION 4.9: The proof relies mainly on a localization argu-
ment and Lemma 4.10. If n =

c and ,
n
is as in Lemma 4.10, we show next that
the function ,
n
is almost orthogonal to e
i
y
( x

1
if c is sufciently small. We
consider, for example, a function

1
3,1,
of the form
=
2
(
Q
C
0
t)
}D
2
(
Q
C
0
t)
b
}
cos(
}
s)7
y
j
(
k.) i sin(
}
s)W
y
j
(
k.)
.
for some arbitrary coefcients (
b
}
)
}
, and we also set
=
2
(
Q
C
0
t)
}D
2
(
Q
C
0
t)
b
}
cos(
}
s)7
(tx)
(k( s).) i sin(
}
s)W
(tx)
(k( s).)
.
We are going to evaluate the real part of the integral
1,
"
e
i
y
( x
,
n

.
First of all, since [k( s)

k[ _ Cn and [
}
( s)[ _ C(n
2
) on the support of
,
n
, we have that
|7
(tx)
(k( s).) 7
y
j
(
k.)|
1
2
(R
n1
)
= O(n
2
) in supp(,
n
).
so, clearly
(4.55) m
1,
"
e
i
y
( x
,
n

=
m
1,
"
e
i
y
( x
,
n

O(n
2
)|,
n
|
1
2
(1,
"
)
| |
1
2
(1,
"
)
.
Next we write
n(s) = ,
n
(s)
2
(
Q
C
0
t)
}D
2
(
Q
C
0
t)
b
}
sin(
}
s).
and notice that
n
0
(s) = ,
0
n
(s)
2
(
Q
C
0
t)
}D
2
(
Q
C
0
t)
b
}
sin(
}
s) ,
n
(s)
2
(
Q
C
0
t)
}D
2
(
Q
C
0
t)
b
}

}
cos(
}
s)
= k,
n
(s)
2
(
Q
C
0
t)
}D
2
(
Q
C
0
t)
b
}
cos(
}
s) ,
0
n
(s)
2
(
Q
C
0
t)
}D
2
(
Q
C
0
t)
b
}
sin(
}
s) (4.56)
,
n
(s)
2
(
Q
C
0
t)
}D
2
(
Q
C
0
t)
b
}
(k
}
) sin(
}
s).
Using this formula and the same argument as for (4.55), we get (recall our notation
before (4.45))
(4.57) ,
n
=
1
k
n
0
(s)7
i n(s)W
O(n
2
)|,
n
|
1
2
(1,
"
)
.
Notice that, by the explicit form of n and
}
, for any integer m we have
| n|
2
1
m
(0,1t])
_ C
n
| n|
2
1
2
(0,1t])
:
therefore, if we write n with respect to the basis
/
as (notice the shift of index
before (4.42))
(4.58) n(s) =
1
/D}
"
`
b
/
/
(cs).
we also nd that
1
/D}
"
(1 [v
/
[)
n
`
b
2
/
_ C
n
| n|
2
1
m
(0,1t])
_ C
n
| n|
2
1
2
(0,1t])
_ C
n
| |
2
1
2
(1,
"
)
.
(4.59)
Differentiating (4.58) with respect to s and using the denition of
}
together with
(4.43), we nd that
n
0
(s) =
1
/D}
"
`
b
/
c
0
/
(cs) =
1
/D}
"
`
b
/
(k O(v
/
))
/
(cs).
The last formula and (4.57) imply
(4.60)
,
n
=
1
/D}
"
`
b
/
(
/
7
i
/
W
/D}
"
`
b
/
O(v
/
)
/
7
O(n
2
)|,
n
|
1
2
(1,
"
)
=
1
/D}
"
`
b
/
3,/

1
/D}
"
`
b
/
(
3,/

/
7
i
/
W
/D}
"
`
b
/
O(v
/
)
/
7
O(n
2
)|,
n
|
1
2
(1,
"
)
.
In the support of ,
n
there exists

0 R such that (}(cs))c =

} s

0 O(n),
which yields
1,
"
e
i
y
( x
,
n

=
1,
"
e
i
f."s/
"
,
n

O(n)||
1
2
(supp(;
))
| |
1
2
(1,
"
)
.
Now, recalling that n =

c and that and e
i((tx)t
1
are orthogonal, from the

last two formulas we obtain
1,
"
e
i
y
( x
,
n

=
1

2

3
O(n
2
)||
1
2
(supp(;
))
| |
1
2
(1,
"
)
(4.61)
where
1
=
supp(;
)
e
i
y
( x
j/j
2
"
`
b
/
3,/
.
2
=
supp(;
)
e
i
y
( x
/D}
"
`
b
/
(
3,/

/
7
i
/
W
).
3
=
supp(;
)
e
i
y
( x
/D}
"
`
b
/
O(v
/
)
/
7
.
To estimate these terms we rst notice that, by the normalization of
}
before
(4.43), the coefcients
`
b
/
in (4.58) can be computed as
`
b
/
= c
1t
0
n(s)
1 2}
0
O
3;
k
/
(cs)Js
= c
2
(
Q
C
0
t)
}D
2
(
Q
C
0
t)
b
}
1t
0
,
n
(s)
sin(
}
s)
1 2}
0
O
3;
k
/
(cs)Js.
Using this formula, Lemma 4.10, and the Hlder inequality, we nd that for any
integer m
(4.62)
`
b
2
/
_ C
n
c
2nC2

2
(
Q
C
0
t)
}D
2
(
Q
C
0
t)
[
b
}
[
2
_ C
n
c
2nC1
2
(
Q
C
0
t)
}D
2
(
Q
C
0
t)
b
2
}
_ C
n
c
2nC2
| |
2
1
2
(1,
"
)
. [l[ _

2
c
.
From the explicit expression for the functions
3,
the above term
1
can be esti-
mated as
1
_ C
1
c
j/j
2
t
(1 [v
/
[
2
)
2
`
b
2
/
1
2
||
1
2
(supp(;
))
.
As before, the factor
1
t
inside the brackets arises from the fact that we are integrat-
ing over 0.
1
t
|.
Using the fact that C
1
[cl[ _ [v
/
[ _ C([cl[ c
2
l
2
) for [l[ _
2
c (which
follows from Weyls asymptotic formula for the eigenvalue problem in (4.42)),
(4.59), and (4.62), we nd that for any large integer m and any J (1.
n
S
)
2
"
j/jt
d
(1 [v
/
[
2
)
2
`
b
2
/
_ C
n
c
11C2nSd
| |
2
1
2
(1,
"
)
.
j/jt
d
(1 [v
/
[
2
)
2
`
b
2
/
_ C
n
c
(d1)(n4)
| |
2
1
2
(1,
"
)
.
By the arbitrariness of m it follows that for any m
0
N
[
1
[ _ C
n
0 c
n
0
| |
1
2
(1,
"
)
||
1
2
(supp(;
))
.
Dividing the set of indices l into {[l[ _
2
c] and {[l[ _
2
c] and using similar
arguments (also taking into account (4.46) and (4.47)), we get
[
2
[ [
3
[ _ C||
1
2
(supp(;
))
1
c
1
/D}
"
(v
2
/
v
4
/
)
`
b
2
/
1
2
_ C
2
||
1
2
(supp(;
))
| |
1
2
(1,
"
)
.
Therefore, using (4.55) and (4.61) we nd
m
1,
"
e
i
y
( x
,
n

= O(n
2
)||
1
2
(supp(;
))
| |
1
2
(1,
"
)
.

1
3,1,
.
Similar estimates hold for span{

1
1,
.

1
2,
.

1
3,2,
], so we obtain
(4.63)
1,
"
e
i
y
( x
,
n

=
O(
2
n)||
1
2
(supp(;
))
| |
1
2
(1,
"
)
for every

1
.
Next we let

1
t
denote the operator in (4.26) with coefcients frozen at s. Since
e
i
y
( x

1
consist of all the eigenfunctions of

1
t
(up to an error o(
2
)) with eigen-
values smaller in absolute value than
2
(see Proposition 4.5 and Remark 4.8), from
(4.63) we then deduce
(4.64) |
1
t
(e
i
y
( x
,
n
)|
1
2
(1,
"
)
_
2
C
|,
n
|
1
2
(1,
"
)
O(
4
2
n)||
1
2
(supp(;
))
for some xed constant C independent of .
It is now possible to choose the cutoff function , (see the comments before
Lemma 4.10) so that it is even and compactly supported in 2. 2|, , 1 in
1. 1|, and ,(2 t ) ,(t ) 1 for t 1. 2|. With this choice, we can nd a
partition of unity (,
n,}
)
}
of 0.
1
t
| consisting of translates of ,
n
(plus a negligible
scaling), with ; running between 1 and a number of order 1
c. For each index

; we choose a point s
}
in the support of ,
n,}
, and we denote by

1
}
the operator
corresponding to (4.26) with coefcients frozen at s
}
. Then, using (4.64), with
easy computations we nd
|1
1
t
|
2
1
2
(1,
"
)
=
1
1
t
}
,
n,}
2
1
2
(1,
"
)
=
1
}
(,
n,}
)
2
1
2
(1,
"
)
O(
c)||
2
1
2
(1,
"
)
_

4
C
,
n,}
2
1
2
(1,
"
)
O(
c)||
2
1
2
(1,
"
)
=

4
C
||
2
1
2
(1,
"
)
O(
c)||
2
1
2
(1,
"
)
for some C independent of . To complete the proof, we need to bound from below
the norm of
t
1
1
t
, showing that
|
t
1
1
t
|
2
1
2
(1,
"
)
_

2
C
||
2
1
2
(1,
"
)
.
To see this, by the last formula it is sufcient to have
(4.65) (1
1
t
. e
i
f."s/
"
)
1
2
(1,
"
)
=
o(
2
)||
1
2
(1,
"
)
||
1
2
(1,
"
)
for any 1
.
We prove this claim for 1
1,
only; for the other 1
},
s the arguments are
similar (see Remark 4.11 below for more details). Setting =
i
i
i
we nd
(see (1.17))
1
1
t
if."s/
"
= e
if."s/
"
(L
i
i
i L
i
i
) e
if."s/
"
d
2
i
ds
2
i
d
2
i
ds
2
2i}
0
e
if."s/
"
d
i
ds
i
d
i
ds
.
When differentiating with respect to s, we either hit the functions c
}
s (and their
derivatives) or other functions like k(cs) or }
0
(cs) (see the denition of 1
1,
above). The latter ones have a slow dependence in s, and therefore these terms
can be collected within an error of the form O(c)||
1
2
(1,
"
)
.
However, by our choices of the second and third parts of the elements in 1
1,
(see Remark 4.8, in particular formulas (4.31), (4.32), and (4.33)), terms containing
zeroth- or rst-order derivatives of c
}
will have coefcients bounded by c, while the
only term containing second derivatives of c
}
will be a linear combination (in ; )
of the expressions
c
2
h(cs)
pC1
4
1
2A
2
0
] 1
h( s)
c0
'c
00
}
(cs). V
+
U(k( s).)). ; = 0. . . . .
c
.
The remaining terms will contain third and the fourth derivatives of c
}
only (mul-
tiplied, respectively, by c
3
and c
4
). Therefore, if we set
1,}
= h(cs)
pC1
4
'c
}
(cs). V
+
U(k.))
i c'c
0
}
(cs). .)
}
0
k
U(k.)
c
2
k
2
'c
00
}
(cs). V(k.))
.
by the above comments and the fact that Jc
}
= h(cs)
0
z
}
c
}
(see (4.38)) we have
1
1
t
e
i
f."s/
"
t
}D0
a
}
1,}
= e
i
f."s/
"
t
}D0
z
}
a
}
1,}
1().
=
t
}D0
a
}
1,}
.
(4.66)
where 1() contains terms of order c or linear combinations of third and fourth
derivatives of c
}
(cs). Thus, using Fourier analysis, we can derive the estimate
|1()|
1
2
(1,
"
)
_ C
1
c
t
}D0
a
2
}
(c c
3
;
3
)
2
1
2
_ C(c
3
)||
1
2
(1,
"
)
(4.67)
for some constant C > 0. Therefore, using (4.66) and (4.67), we obtain
1
1
t
. e
i
f."s/
"
1
2
(1,
"
)
= O(c
3
)||
1
2
(1,
"
)
||
1
2
(1,
"
)
.
which yields (4.65) and concludes the proof.
Remark 4.11. The last step in the proof of Proposition 4.9 is nearly identical for
1
2,
except that, still by the computations in Remark 4.8, in the counterpart of
(4.66) we will obtain j
}
instead of z
}
(see (4.38)). When considering 1
3,
, setting
=

2
t
}D
2
t
a
}
3,}
(see (4.48)), we nd
1
1
t
e
i
f."s/
"
= e
i
f."s/
"
2
t
}D
2
t
v
}
a
}
3,}

1().
|

1()|
1
2 _ C
1
c
2
t
}D
2
t
a
2
}
(c c
2
;
2
)
2
1
2
_ C(c
4
)||
1
2
(1,
"
)
.
(4.68)
4.5 Invertibility of L
"
in Weighed Spaces
Our goal is to show that the linearized operator 1
t
(see (4.7)) at approximate
solutions is invertible on spaces of functions satisfying suitable constraints. We
begin with some preliminary notation and lemmas; we rst collect decay properties
of Greens kernels in Euclidean space. Let us consider the equation
(4.69) û u = } in R
n1
.
where } decays to 0 at innity. The solution of the above equation can be repre-
sented as
u(.) =
R
n1
G
0
([. .[)}(.)J..
where G
0
: R
C
R
C
is a function singular at 0 that decays exponentially to 0 at
innity. Using the notation of Subsection 4.2 and standard elliptic regularity theory,
we can prove the following result (the choice _
1
2
for the Hlder exponent is
technical and is used in the proof of Lemma 4.13).
LEMMA 4.12 Let

O > 0, _
1
2
, 0 < t < 1, 0 < c < 1, and } C
r
.
Then equation (4.69) has a (unique) solution u of class C
2,
r
that vanishes on
dT
N
O
(0). Moreover, there exist c
0
> 0 sufciently close to 1 and
`
C
0
sufciently
large (depending only on n, , min{

O. 1], and c) such that for c
0
_ c < 1
|u|
C
2;
&
_
`
C
0
|} |
C
&
.
Let now t. c (0. 1) (to be xed later). For any integer m we let

C
n,r
,
denote
the weighted Hlder space
(4.70)

C
n,r
=

u : R
n1
C : sup
+2R
n1
e
j+j
|u|
C
m;
(B
1
(+))
< o
.
We also consider the following set of functions
1
t
-periodic in s:
1
2
(

C
n,r
) =

u : 0.
1
t
| R
n1
C :
s u(s. ) 1
2
0.
1
t
|:

C
n,r
.
(4.71)
and for l N, we similarly dene the functional space
H
/
(

C
n,r
) =

u : 0.
1
t
| R
n1
C :
s u(s. ) H
/
(0.
1
t
|: C
n,r
.
(4.72)
The weights here are suited for studying functions that decay in . like e
j+j
, as
the fundamental solution of ^
R
n1u u = 0. The parameter c < 1 has been
introduced to allow some exibility in the decay rate. When dealing with functions
belonging to the above three spaces, the symbols
| |
N
C
m;
&
. | |
N
1
2
(
N
C
m;
&
)
. and | |
N
1
l
(
N
C
m;
&
)
will denote norms induced by formulas (4.70), (4.71), and (4.72). Also, we keep
the same notation for the norms when considering functions dened on subsets of
0.
1
t
| R
n1
.
We next consider some positive constants

V .

} .

h, and

k that satisfy the relations
in (1.9). If and

1
are as in the previous subsection and

as in Section 4, letting
D
1,t
= 0.
1
t
| T
t
C1
(0) _ 0.
1
t
| R
n1
.
we dene the space of functions
H
t
:=
: m
T
L;"
(s. .)e
i
y
( x
s. .
= 0 for all

1
.
This condition represents, basically, orthogonality with respect to

1
(multiplied
by the phase factor), when the function is scaled in . by

V . This is a choice
made for technical reasons, which will be helpful in Proposition 4.14. We next
have the following result, related to Proposition 4.9 once we scale ..
LEMMA 4.13 Let
1
2
_ t < 1 and c (0. 1). Then, for small there exists
a positive constant C, depending only on ], t, c, 1,

V , and

} , such that the
following property holds: for

} small, for c 0, and for any function b

1
2
(

C
r
)
there exist u

H
t
and

1
such that, in D
1,t
,
(4.73)
1
y
V
d
2
xx
u ^
+
u u
y
h
p1
y
V
U(.
V )
;1
u
(] 1)
y
h
p1
y
V
U(.
V )
;1
e
i
y
( x
m(e
i
y
( x
u) = b e
i
y
( x
.
u = 0 on dD
1,t
(notice that above is intended to be scaled in .) and such that we have the
estimates
|u|
N
1
2
(
N
C
2;
&
)
|u|
N
1
1
(
N
C
1;
&
)
|u|
N
1
2
(
N
C
&
)
_
C
2
inf
2
y
1
|b e
i
y
( x
|
N
1
2
(
N
C
&
)
.
(4.74)
||
N
1
2
(
N
C
&
)
_ C|b|
N
1
2
(
N
C
&
)
. (4.75)
PROOF: First of all we observe that a solution to (4.73) of class 1
2
exists. In
fact, by replacing D
1,t
with 0.
1
t
| R
n1
, this would simply follow from Propo-
sition 4.9 with V

V . However, since the functions in

1
decay exponentially
to 0 as [.[ o, the Dirichlet boundary conditions do not affect the solvability
property; for more details, see, for example, [40, lemma 5.5]. Notice that indeed,
by (1.18) and Proposition 4.5(v), the elements of 1
decay at the rate e
y
kj+j
, and
by (1.9),

k >
V . In particular, ||
N
1
2
(
N
C
&
)
is nite and (4.75) holds. We also
have (4.74) replacing the left-hand side by the 1
2
norm of u. We divide the rest of
the proof into two steps.
Step 1. u

1
2
(

C
r
) AND |u|
N
1
2
(
N
C
&
)
_
C
2
|b|
N
1
2
(
N
C
&
)
. We set u = e
i
y
( x
and c = e
i
y
( x
(b ), so satises
^

1
y
(
2
y
V
2i
y
(
y
V
d
x

y
h
p1
y
V
U(.
V )
;1
(] 1)
y
h
p1
y
V
U(.
V )
;1
m( ) = c in T
t
C1
(0).
= 0 on dT
t
C1
(0).
We now use a Fourier decomposition in the variable s; setting
c(s. .) =
}
c
}
(.)e
i}tx
. (s. .) =
}
(.)e
i}tx
(here we are assuming for simplicity that 1 = 2), we see that each c
}
belongs to
C
r
,
y
V
, that
}
|c
}
|
2
N
C
&
=
1
c
|c|
2
N
1
2
(
N
C
&
)
.
}
|
}
|
2
N
C
&
=
1
c
||
2
N
1
2
(
N
C
&
)
_
C
c
2
|b|
N
1
2
(
N
C
&
)
.
(4.76)
and that each
}
solves
(4.77)
^
+
}

1
y
(
2
Ct
2
}
2
2
y
( t}
y
V
y
h
p1
y
V
U(.
V )
;1
}
(] 1)
y
h
p1
y
V
U(.
V )
;1
m(
}
) = c
}
in T
t
C1
(0).
}
= 0 on dT
t
C1
(0).
From elliptic regularity theory, we nd that for any 1 > 0 there exists a constant
C depending only on 1, ], and t such that
|
}
|
C
(B
R
)
_ C|c
}
|
N
C
&
C|
}
|
1
2.
Now we choose 1 (depending on ] and c) so large that
]
h
;1
V
U
;1
<
1
4
(1 c) for [.[ _
1
2
.
and a smooth radial cutoff function , such that ,(.) = 1 for [.[ _
T
2
, and
,(.) = 0 for [.[ _ 1. Next, we write equation (4.77) as
^
+
}

1
y
(
2
Ct
2
}
2
2
y
( t}
y
V
}
(1 ,)]
y
h
p1
y
V
U(.
V )
;1
}
=
c
}
,]
y
h
p1
y
V
U(.
V )
;1
}
.
}
= 0 on dT
t
C1
(0).
We notice that the rst linear coefcient of
}
is bounded below (uniformly in ; )
by 1. Therefore, using Greens representation formula, the maximum principle,
and our choice of 1 (see Lemma 4.12) for any c
0
< c, we have the estimate
|
}
|
N
C
&
0
_ C(|c
}
|
N
C
&
|
}
|
1
2)
for some xed constant C depending only on ], c, and t. Taking the square and
summing over ; , we get
|u|
2
N
1
2
(
N
C
&
0
)
= ||
2
N
1
2
(
N
C
&
0
)
_ C|b|
2
N
1
2
(
N
C
&
)
C||
2
1
2
_ |b|
2
N
1
2
(
N
C
&
)
.
We next want to replace in the last formula c
0
with c. Rewrite (4.73) as
1
y
V
d
2
xx
u ^
+
u u =c :=
y
h
p1
y
V
U(.
V )
;1
u
(] 1)
y
h
p1
y
V
U(.
V )
;1
e
i
y
( x
m(e
i
y
( x
u) b e
i
y
( x
.
u = 0 on dD
1,t
.
Using the same procedure as above, write c(s. .) =

}
c
}
(.)e
i}tx
and u(s. .) =
}
u
}
(.)e
i}tx
.
We now consider the function U
;1
u
}
: by (1.18), if we choose
c
0

(] 1)
V
> c.
it follows from the above estimates that |c|
N
1
2
(
N
C
&
)
is nite and that
}
|c
}
|
2
N
C
&
_
C
c
|b|
2
N
1
2
(
N
C
&
)
.
Moreover, u
}
satises
û
}

1
t
2
}
2
y
V
u
}
=c
}
in T
t
C1
(0).
u
}
= 0 on dT
t
C1
(0).
Also, it is easy to show
(4.78)
|u|
2
N
1
2
(
N
C
2;
&
)
|u|
2
N
1
1
(
N
C
1;
&
)
|u|
2
N
1
2
(
N
C
&
)
=
1
c
|u
}
|
2
N
C
2;
&
(1 c
2
;
2
)|u
}
|
2
N
C
1;
&
(1 c
2
;
2
c
4
;
4
)|u
}
|
2
N
C
&
.
and therefore we are reduced to nding the estimates
|u
}
|
2
N
C
2;
&
. |u
}
|
2
N
C
1;
&
. |u
}
|
2
N
C
&
.
done in the next step.
Step 2. CONCLUSION OF PROOF. We now set a
}
= 1 c
2
;
2
V and
}
(.) = u
}
(.
a
}
). Then, from a change of variables we have the equation
^
}
(.)
}
(.) =

J
}
(.) :=
1
o
j
c
}
+
p
o
j
in Tp
o
j
(t
C1)
(0).
}
= 0 on dTp
o
j
(t
C1)
(0).
Notice that a
}
> 0 stays bounded from below independently of ; , and therefore
by a scaling argument (and some elementary inequalities), we nd
(4.79)
sup
+,:2B
1
(x)
[

J
}
(.)

J
}
(z)[
=
1
a
}
sup
+,:2B
1=
p
a
j
(x)
c
}
a
}
c
}
a
}
_
C
a
}
sup
+,:2B
1
(x)
c
}
a
}
c
}
a
}
_
C|c
}
|
N
C
&
a
1C
2
}
sup
+,:2B
1
(x)
[. z[
r
e
&jxj
p
a
j
where C depends only on t, and hence we get
(4.80) |

J
}
|
N
C
&=
p
a
j
_
C
a
1Cr2
}
|c
}
|
N
C
&
.
Now Lemma 4.12 implies that
|
}
|
N
C
2;
&=
p
a
j
_
C
a
1Cr2
}
|c
}
|
N
C
&
.
Fromthis estimate, we will obtain some control on u
}
by scaling back the variables.
We consider an arbitrary . R
n1
. As before, we have
sup
+,:2B
1
(x)
[u
}
(.) u
}
(z)[
[. z[
r
= sup
+,:2B
1
(x)
[
}
(
a
}
.)
}
(
a
}
z)[
[. z[
r
.
Since a
}
can be arbitrarily large, we cannot evaluate the difference
}
(
a
}
.)
}
(
a
}
z) directly using the weighted norm in the denition (4.70) (as we did for
the rst inequality in (4.79)), since the two points

a
}
. and

a
}
z might not be-
long to the same unit ball. We avoid this problem by choosing
a
}
| (the integer
part of

a
}
) points (.
/
)
/
lying on the segment
a
}
..
a
}
z| at equal distance
from each other and using the triangle inequality. Now the distance of two consec-
utive points .
/
and .
/C1
will stay uniformly bounded from above, and the minimal
norm of the .
/
s is bounded from below by C
1
a
}
([.[ 1). Therefore, adding
a
}
| times the inequality and using (4.80), we obtain
sup
+,:2B
1
(x)
[u
}
(.) u
}
(z)[
[. z[
r
_
C
a
}
[. z[
r
C
a
1Cr2
}
. z
a
}
r
e
p
o
j
.jxj1/
p
a
j
|c
}
|
N
C
&
_
C
a
1Cr12
}
e
jxj
|c
}
|
N
C
&
_
C
a
}
e
jxj
|c
}
|
N
C
&
_
C
a
}
e
jxj
| c
}
|
N
C
&
.
since we chose t _
1
2
and since a
}
is uniformly bounded from below.
Similarly, taking rst and second derivatives, we nd that
sup
+,:2B
1
(x)
[Vu
}
(.) Vu
}
(z)[
[. z[
r
_
C
a
}
e
jxj
| c
}
|
N
C
&
.
sup
+,:2B
1
(x)
[V
2
u
}
(.) V
2
u
}
(z)[
[. z[
r
_ Ce
jxj
| c
}
|
N
C
&
.
where, again, C depends only on t. Recalling that a
}
=

V c
2
;
2
, we have in
this way proved that
|u
}
|
2
N
C
2;
&
_ C| c
}
|
2
N
C
&
. |u
}
|
2
N
C
1;
&
_
C
1 c
2
;
2
| c
}
|
2
N
C
&
.
|u
}
|
2
N
C
&
_
C
(1 c
2
;
2
)
2
| c
}
|
2
N
C
&
.
Now the conclusion follows from (4.76), (4.78), the last formula, and the fact that
| c|
N
1
2
(
N
C
&
)
_ C inf
2
y
1
|b |
N
1
2
(
N
C
&
)
:
see the beginning of Step 1.
We next consider the operator 1
t
in

D
t
(see (4.7)), acting on a suitable subset
of H
Q
T
"
(verifying an orthogonality condition similar to (4.51)). We want to allow
some exibility in the choice of approximate solutions: to do this we consider a
normal section to , that satises the following two conditions:
(4.81) span
h
pC1
4
c
}
: ; = 0. . . . .
and ||
1
2
(0,1)
_ c
1
c.
Here (c
}
)
}
are as in (4.38), while c
1
is a large constant to be determined later.
Notice that by (4.81) has, naively, only Fourier modes with index bounded by
1
t
. This yields estimates of the type |
(kC1)
|
1
2 _ C
k,
1
e
|
(k)
|
1
2. We have
therefore |
000
|
1
2
0,1]
_ C, which implies |
00
|
1
1 _ C, so also (4.4) holds
true. This will allow us, in the next section, to apply Proposition 4.1.
Next, consider the variables z dened in (3.1). In the coordinates (s. z), we will
consider the approximate solution
(4.82)

[
t
= e
i
Q
f ."s/
"
n
t
h(cs)U(k(cs)z) U
1
(s. z)
:=

[
0,t

[
t
.
where cs = s, and where

} and U
1
satisfy, for some xed C > 0 and t (0. 1),
|

} } |
1
2
(0,1])
_ Cc
2
. [U
1
[(s. z) _ Cc(1 [z[
C
)e
k(N x)j:j
. (4.83)
|[hU(k) U
1
[
;1
[hU(k)[
;1
|
C
_ Cc in

D
t
. (4.84)
With this choice of

[
t
, we are going to study the analogue of Lemma 4.13 for 1
t
(see (4.7)), using a perturbation method.
To state our nal result, we need to introduce some more notation. Recalling the
denition in (4.70) and still using the coordinates (s. z) for t (0. 1) and c > 0,
we dene the function space
(4.85) 1
2
(C
n,r
,V
) =
u :

D
t
C : s u
s.
V(cs)
1
2
0.
1
t
|: C
n,r
,1
.
Also, for m N, we similarly dene
(4.86) H
/
(C
n,r
,V
) =
u :

D
t
C : s u
s.
V(cs)
H
/
0.
1
t
|: C
n,r
,1
.
We next let

1
be the counterpart of 1
(see (4.50)), when we replace the

coordinates . by z. Finally, we denote by

H
t
the following subspace of functions:
(4.87)

H
t
:=
H
Q
T
"
: m
Q
T
"
e
i
Q
f ."s/
"

= 0 for all

1
.
Dening
(4.88) | |
,V
:= | |
1
2
(C
2;
&;V
)
| |
1
1
(C
1;
&;V
)
| |
1
2
(C
&;V
)
.
we then have the following result (recall the denition of

D
t
in (4.10)):
PROPOSITION 4.14 Suppose 0 < c < 1 and
1
2
_ t < 1. Suppose

[
t
is as in
(4.82), with

} and U
1
satisfying (4.83). Then, if 1
2
(cs) = V(cs) and if A and
are sufciently small, in the limit c 0 the following property holds: for any
function b 1
2
(C
r
,V
) there exist u

H
t
and

1
such that
(4.89)
^
v
"
u V(c.) u [

[
t
[
;1
u
(] 1)[

[
t
[
;3

[
t
m(

[
t
u) = b e
i
Q
f ."s/
"
in

D
t
.
u = 0 on d

D
t
.
is solvable, and such that for every c
0
< c there exists some C > 0 for which we
have the estimates
| u|
0
,V
_
C
2
inf
Q 2
Q
1
|b e
i
Q
f ."s/
"
|
1
2
(C
&;V
)
| |
1
2
(C
&;V
)
_ C|b|
1
2
(C
&;V
)
.
(4.90)
PROOF: We divide the proof into two steps.
Step 1. SOLVABILITY OF (4.89). First of all, we notice that, from Proposi-
tion 4.9 and from elliptic regularity results, if H
t
denotes the subspace of function
in H
2
(N,
t
) satisfying (4.51), then the operator 1
1
t
is invertible from
(H
t
. | |
1
2
(1,
"
)
) onto (
t
1
2
(N,
t
). | |
1
2
(1,
"
)
); moreover, the norm of the
inverse operator is bounded by C
2
.
From the comments at the beginning of the proof of Lemma 4.13, we also de-
duce the following property: Given b 1
2
({[.[ _ (c
1)1(cs)]) there exist

u H
2
({[.[ _ (c
1)1(cs)]) and 1
such that
(4.91)
1
t
u := ^
y v
"
u V(cs)u h(cs)
;1
U(k(cs).)
;1
u
(] 1)h(cs)
;1
U(k(cs).)
;1
e
i
Q
f ."s/
"
m(e
i
Q
f ."s/
"
u)
= b e
i
Q
f ."s/
"
in {[.[ _ (c
1)1].
m
fj+j(t
C1)1g
ue
i
f."s/
"
JV
y v
"
= 0 for every 1
.
Again, we have the estimates
(4.92)
|u|
1
2
(fj+j(t
C1)1g)
_
C
2
|b|
1
2
(fj+j(t
C1)1g)
.
||
1
2
(fj+j(t
C1)1g)
_ C|b|
1
2
(fj+j(t
C1)1g)
.
Using a perturbative argument, we show that we can recover the same invertibility
result for (4.89) where, compared to (4.91), we need to substitute . with z, ^
y v
"
with ^
Q v
"
, } with

} , and e
i(((tx)t
hU with

[
t
.
In fact, let us denote by
+
and
:
the orthogonal projections in 1
2
onto the
orthogonal complements of the sets {e
i(((tx))t
: 1
], {e
i
Q
( (tx)t
:

1
] with respect to the scalar products induced by the metrics g

t
and g
t
,
respectively. By (4.81), Lemma 3.1, and (4.83) for every u H
2
(

D
t
) and every
b 1
2
(

D
t
), we have
|1
t
u

1
t
u|
1
2
(
Q
T
"
)
_ C(c
1
)c|u|
1
2
(
Q
T
"
)
.
|
+
b
:
b|
1
2
(
Q
T
"
)
_ C(c
1
)c|b|
1
2
(
Q
T
"
)
.
where C(c
1
) is a positive constant which depends on ,, V , and the constant c
1
in
(4.81).
From (4.92) and the last formula we deduce the solvability of (4.89), together
with the estimates | u|
1
2
(
Q
T
"
)
_ (C
2
)|b|
1
2
(
Q
T
"
)
and | |
1
2
(
Q
T
"
)
_ C|b|
1
2
(
Q
T
"
)
.
Step 2. PROOF OF (4.90). Recall that the coordinates . (see the beginning of
this section) are not global, since they are dened locally in s by normal parallel
transport; the same holds, of course, for the coordinates z. Therefore, if we prolong
the zs along ,
t
, there will be a discontinuity between 0 and
1
t
.
To reduce ourselves to the periodic case, as in Lemma 4.13, we apply a rotation
1
t
= 1
t
(cs) to the z-axes that makes the coordinates z := 1(cs)z periodic in s.
To compute the Laplace-Beltrami operator in the new coordinates z, we should
apply the chain rule in this way:
d
:
j
u = (1
t
)
}/
d
Q :
l
u. d
2
x:
j
u = cd
N x
(1
t
)
}/
d
Q :
l
u (1
t
)
}/
d
2
x Q :
l
u.
d
2
:
j
:
l
u = 1
n}
1
t/
d
2
Q :
m
Q :
t
u.
In particular, since 1
t
is orthogonal,
d
2
:
j
:
j
u = 1
n}
1
t}
d
2
Q :
m
Q :
t
u = (1
t
)
n}
(1
1
t
)
}t
d
2
Q :
m
Q :
t
u
= d
2
Q :
m
Q :
m
u.
namely, the main term in the Laplacian stays invariant. Taking into account Lem-
ma 3.1 and the last formulas, for c
00
(c
0
. c) we nd
(4.93) |^
Q :
Q v
"
u ^
:
Q v
"
u|
1
2
(C
&
00
;k
)
_ C(c
1
)c|u|
00
,V
.
We next use a localization argument as in the proof of Proposition 4.9. If s
}
and
,
n,}
are as in that proof, by (4.83) we can nd

0
}
R such that

} (cs)c

}
}
s
0
}
= O(
c) in the support of ,
n,}
. If we set ^
(x,Q :)
R
n
= ^
Q :
R
n1
d
2
xx
, and if we
scale the z-variables by 1(cs) =
V(cs), the function ,

n,}
(s)u(s. z) (which is
now periodic in s) satises the equation
1
y
V
j
d
2
xx
,
n,}
u ^
Q :
R
n1
(,
n,}
u) ,
n,}
u
y
h
p1
j
y
V
j
U
k
}
V
}
;1
,
n,}
u
(] 1)
y
h
p1
j
y
V
j
U
k
}
V
}
;1
e
i(
y
( xC
y
0
j
)
m(e
i(
y
( xC
y
0
j
)
,
n,}
u)
= F
}
in

[ z[ _
t
C1
1
.
,
n,}
u = 0 on

[ z[ =
t
C1
1
.
where
F
}
=
1
V( s)
,
n,}
e
i
Q
f ."s/
"
(b )
1
V( s)
^
Q :
Q v
"
^
(x,Q :)
R
n
,
n,}
u
1
V( s)
2V
Q :
Q v
"
u V
Q :
Q v
"
,
n,}
u^
Q :
Q v
"
,
n,}

1
V( s)
(
V
}
V ),
n,}
u
1
V( s)
[

[
t
[
;1
,
n,}
u
1
V( s)
(] 1)[

[
t
[
;2
[
t
m(

[
t
,
n,}
u)
h
;1
}
V
}
U
k
}
V
}
;1
,
n,}
u
(] 1)
h
;1
}
V
}
U
k
}
V
}
;1
e
i(
y
( xC
y
0
j
)
m(e
i(
y
( xC
y
0
j
)
,
n,}
u).
In the last formula, the functions b, , V , and

[
t
are intended to be scaled in z
by

V(cs). Reasoning as for (4.63), from (4.87) we nd that
Q
T
"
e
i
y
(
j
x
y
0
j
,
n,}

= O(
2

c)||
1
2
(supp(;
;j
))
| |
1
2
(
Q
T
"
)
for every

1
. Moreover, as for (4.93) we can show that

|(^
Q :
Q v
"
^
(x,Q :)
R
n
),
n,}
u|
N
1
2
(
N
C
&
00
)
_
C(c
1
)c
|,
n,}
u|
N
1
2
(
N
C
2;
&
00
)
|,
n,}
u|
N
1
1
(
N
C
1;
&
00
)
|,
n,}
u|
N
1
2
(
N
C
&
00
)
.
Therefore, using Lemma 4.13, (4.83), and (4.84) we obtain the estimate
|,
n,}
u|
N
1
2
(
N
C
2;
&
00
)
|,
n,}
u|
N
1
1
(
N
C
1;
&
00
)
|,
n,}
u|
N
1
2
(
N
C
&
00
)
_
C
2
|F
}
|
N
1
2
(
N
C
&
00
)
_
C
2
|,
n,}
b|
N
1
2
(
N
C
&
00
)

C
2
|,
n,}
|
N
1
2
(
N
C
&
00
)
C
|u|
N
1
2
(
N
C
2;
&
00
,supp(;
;j
))
|u|
N
1
1
(
N
C
1;
&
00
,supp(;
;j
))
|u|
N
1
2
(
N
C
&
00
,supp(;
;j
))
.
(4.94)
where the last symbols denote the restrictions of the weighted norms to supp(,
n,}
).
Recall that the functions in the previous formula have been scaled in z by

V(cs);
therefore, from the uniform continuity of V( s), for some C > 0 we have (recall
that c
00
(c
0
. c))
1
C
|,
n,}
u|
0
,V
_ |,
n,}
u( .
V( s) )|
N
1
2
(
N
C
2;
&
00
)
|,
n,}
u( .
V( s) )|
N
1
1
(
N
C
1;
&
00
)
|,
n,}
u( .
V( s) )|
N
1
2
(
N
C
&
00
)
.
A similar inequality holds for the restriction of u to the support of ,
n,}
, together
with
|,
n,}
b( .
V( s) )|
N
1
2
(
N
C
&
00
)
|,
n,}
( .
V( s) )|
N
1
2
(
N
C
&
00
)
_
C
|,
n,}
b|
1
2
(C
&;V
)
|,
n,}
|
1
2
(C
&;V
)
.
Using the last two inequalities, taking the square of (4.94), and summing over ; ,
we can bring the last term in the right-hand side to the left, so we get (4.90).
5 Approximate Solutions
In this section we construct some approximate solutions to (1.16) that depend
on suitable parameters and nd estimates on the error terms. As in the previous
subsection, we let . be a system of Fermi coordinates in N,
t
, and for a normal
section of N,
t
of class H
2
we dene the coordinates (see (3.1))
z = . (cs). z R
n1
.
By the results in Subsection 4.2, we will restrict our attention to the set

D
t
.
Remark 5.1. In the spirit of Proposition 4.4, we will work with approximate solu-
tions

[
t
supported in

D
t
. Therefore, using the above coordinates,

[
t
(s. z) has to
vanish for [z[ sufciently large. This can be achieved by formally dening

[
t
(s. z)
on N,
t
and multiplying it by a cutoff function n
t
as in Subsection 4.2. However,
since the functions we are dealing with decay exponentially to 0 as [z[ o, the
effect of this cutoff on the expansions below is exponentially small in c, and it will
turn out to be negligible for our purposes. Therefore, for reasons of brevity and
clarity, we will tacitly assume that

[
t
(s. z) is multiplied by such a cutoff without
writing it explicitly.
Recall that in (4.6) we dened
S
t
([) = ^
v
[ V(c.)[ [[[
;1
[.
We set

}
0
( s) = }( s) c}
1
( s), where } is given in (1.12) and }
1
in (3.25) (we
refer to Subsection 2.4 for the denition of A
0
). If n
i
and n
i
are smooth functions
of s, we saw in Subsection 3.2 that
S
t
e
i
Q
f
0
."s/
"
(hU(kz) c(n
i
in
i
))
=
e
i
Q
f
0
."s/
"
(c(R
i
i R
i
)) 1
2
(z)e
k(tx)j:j
for some quantities R
i
and R
i
. We choose n
i
= n
i,e
n
i,o
and n
i
= n
i,e
n
i,o
(see (3.13), (3.9), and (3.12)) to make R
i
and R
i
vanish, using the stationarity
condition (1.15) (see Lemma 3.2).
From Proposition 3.3 we also have that
(5.1)
e
i
Q
f
0
."s/
"
S
t
([
1,t
)
= c
2
(

1
i,e

1
i,o
) c
2
(

1
i,e,(
1

1
i,o,(
1
)
c
2
i(

1
i,e

1
i,o
) c
2
i(

1
i,e,(
1

1
i,o,(
1
) 1
3
(z)e
k(tx)j:j
.
Here we want to prove error estimates when and the phase satisfy some precise
conditions in terms of Fourier analysis and Sobolev norms, and we add further
correction terms.
To allow more exibility in the choice of approximate solutions, we substitute
the phase

}
0
with the function

} = } c}
1
c
2
}
2
, where }
2
is some function
of class H
2
. On and }
2
we assume the following conditions for some constants
c
1
and c
2
to be determined later:
(5.2) ||
1
2 _ c
1
c. |}
2
|
1
2 _ c
2
.
Moreover, letting be as in Subsection 4.3 and c
}
and o
}
as in (4.38), we also
assume that
span
h
pC1
4
c
}
: ; = 0. . . . .
.
}
2
span
h
1
2
o
}
: ; = 0. . . . .
.
(5.3)
Notice that for }
2
we have similar observations to those made for after (4.81).
Since, again, the Fourier modes of }
2
have index bounded by
1
t
, we get estimates
of the type |}
(kC1)
2
|
1
2 _ C
k,
1
e
|}
(k)
2
|
1
2. This allows us to control terms in-
volving }
000
2
, }
(4)
2
, etc. To deal with the resonance phenomenon mentioned in the
introduction, related to the components in 1
3,
of the approximate kernel, we add
to the approximate solutions a function
like
(5.4)
= (cs)7
(tx)
i (cs)W
(tx)
(see (4.37) and the lines after), with and given by
(5.5) (cs) =
2
t
}D
2
t
b
}
}
(cs). =
2
t
}D
2
t
b
}
}
(cs).
where, we recall,
}
solves (4.42) and is related to
}
by (4.43). Below we will
regard as an independent variable, and as a function of . Introducing the
norm
(5.6) ||
r
:=

2
t
}D
2
t
b
2
}
(1 [; [)
2
1
2
.
we will assume later on that
(5.7) ||
r
_ c
3
c
2
for some constant c
3
> 0 to be specied below.
We will look for approximate solutions of the form
2,t
(s. z) := e
i
z
f ."s/
"
h(cs)U(k(cs)z)
cn
i
in
i
| c
2
c
2
.
(5.8)
In this formula

} is as above, while and
0
are corrections whose choice is given
below in order to improve the accuracy of the approximate solutions.
Our goal is to estimate with some accuracy the quantity S
t
(

2,t
): for simplicity,
to treat separately some terms in this expression, we will write

2,t
as
(5.9)

2,t
(s. z) =

1,t
(s. z) 1(s. z) J(s. z) G(s. z).
where

1,t
, 1, J, and G are, respectively, dened by
1,t
(s. z) := e
i
z
f ."s/
"
{h(cs)U(k(cs)z) cn
i
in
i
|]
:= e
i
"
2
f
2
."s//
"
[
1,t
.
1(s. z) := c
2
e
i
z
f ."s/
"
. J(s. z) := c
2
e
i
z
f ."s/
"
0
.
G(s. z) := e
i
z
f ."s/
"
.
To expand S
t
(
2,t
) conveniently, we can write
S
t
(

2,t
) = S
t
(

1,t
) A
3
A
4
A
5
A
6
.
where A
3
. . . . . A
6
are, respectively, the linear terms in the equation that involve 1,
J, and G (see (5.9)):
A
3
= ^
v
1 V(c.)1 [

1,t
[
;1
1
(] 1)[

1,t
[
;3
1,t
m(

1,t

1).
(5.10)
A
4
= ^
v
J V(c.)J [

1,t
[
;1
J
(] 1)[

1,t
[
;3
[
1,t
m(

1,t

J).
(5.11)
A
5
= ^
v
G V(c.)G [

1,t
[
;1
G
(] 1)[

1,t
[
;3
1,t
m(

1,t

G).
(5.12)
and where A
6
contains the contribution of the nonlinear part
(5.13)
A
6
= [

2,t
[
;1
2,t
[

1,t
[
;1
(1 J G)
(] 1)[

1,t
[
;3
1,t
m(

1,t
(

1

J

G)).
Next we write (tautologically)
S
t
(

1,t
) = e
i
"
2
f
2
."s/
"
S
t
(

[
1,t
) A
1
.
A
1
= S
t
(

1,t
) e
i
"
2
f
2
."s/
"
S
t
(

[
1,t
).
(5.14)
and set
(5.15)
A
2
= e
i
Q
f ."s/
"
e
i
Q
f
0
."s/
"
S
t
(

[
1,t
)
c
2
(

1
i,o

1
i,e
) c
2
(

1
i,o,(
1

1
i,e,(
1
)
c
2
i(

1
i,e

1
i,o
) c
2
i(

1
i,e,(
1

1
i,o,(
1
)
.
so that A
2
represents the terms that are formally of order c
3
and higher in S
t
(

[
1,t
)
(multiplied by a phase factor). Therefore, from denitions (5.10)(5.15), we nd
that
(5.16)
S
t
(

2,t
) = e
i
Q
f ."s/
"
c
2
(

1
i,o

1
i,e
) c
2
(

1
i,o,(
1

1
i,e,(
1
)
c
2
i(

1
i,e

1
i,o
) c
2
i(

1
i,e,(
1

1
i,o,(
1
)
A
1
A
2
A
3
A
4
A
5
A
6
.
To estimate rigorously the A
i
s, we display the rst- and second-order deriva-
tives of

2,t
:
d
x

2,t
= i

}
0
(cs)e
i
z
f ."s/
"
h(cs)U(k(cs)z) cn
i
in
i
|
c
2
c
2
e
i
z
f ."s/
"
ch
0
U(kz) chk
0
VU z c
2
d
x
n
i
i c
2
d
x
n
i
c
3
d
x
c
3
d
x
0
d
x
.
d
}

2,t
= e
i
z
f ."s/
"
hkd
}
U(k(cs)z) cd
}
n
i
i d
}
n
i
| c
2
d
}

c
2
d
}
0
d
}
.
d
2
xx

2,t
= (

}
02
i c

}
00
)(cs)e
i
z
f ."s/
"
h(cs)U(k(cs)z) cn
i
in
i
| c
2

c
2
0
a(cs)7(kz)
e
i
z
f ."s/
"
c
2
h
00
U(kz) 2c
2
h
0
k
0
VU z c
2
hk
00
VU z
c
2
hk
02
V
2
Uz. z| c
3
d
2
xx
n
i
i c
3
d
2
xx
n
i
c
4
d
2
xx

c
4
d
2
xx
0
d
2
xx
2i

}
0
(cs)e
i
z
f ."s/
"
ch
0
U(kz) chk
0
VU z c
2
d
x
n
i
i c
2
d
x
n
i
c
3
d
x
c
3
d
x
0
d
x
.
d
2
}/
2,t
= e
i
z
f ."s/
"
h(cs)k
2
d
2
}/
U(k(cs)z) cd
2
}/
n
i
i d
2
}/
n
i
| c
2
d
2
}/

c
2
d
2
}/
0
d
2
}/
.
d
2
x}

2,t
= i

}
0
(cs)e
i
z
f ."s/
"
h(cs)kd
}
U(k(cs)z) cd
}
n
i
i d
}
n
i
| c
2
d
}

c
2
d
}
0
a(cs)kd
}
7(kz)
e
i
z
f ."s/
"
ch
0
kd
}
U(kz) chk
0
d
}
U chkk
0
z
/
d
2
}/
U c
2
d
2
x}
n
i
i c
2
d
2
x}
n
i
c
3
d
2
x}
c
3
d
2
x}
0
d
2
x}
.
To simplify the expressions of the error terms, we introduce some convenient no-
tation. For any positive integer q, the two symbols R
q
(.
0
) and R
q
(.
0
.
00
)
will denote error terms satisfying the following bounds for some xed constants C
and J (which depend on q. c
1
. c
2
, and c
3
but not on c. s, or )
[R
q
(.
0
)[ _ Cc
q
(1 [z[
d
)e
kj:j
.
[R
q
(.
0
) R
q
(
0
)[ _ Cc
q
(1 [z[
d
)[

[ [
0

0
[|e
kj:j
.
while the termR
q
(.
0
.
00
) (which also involves second derivatives of ) stands
for a quantity for which
[R
q
(.
0
.
00
)[ _ Cc
q
(1 [z[
d
)e
kj:j
Cc
qC1
(1 [z[
d
)e
kj:j
[
00
[.
[R
q
(.
0
.
00
) R
q
(
0
.

00
)[
_ Cc
q
(1 [z[
d
)[

[ [
0

0
[|e
kj:j
Cc
qC1
(1 [z[
d
)
[
00

00
[([

[ [
0

0
[)
[
00

00
[
e
kj:j
.
Similarly, we will let R
q
( s) denote a quantity (depending only on s and z) such
that
[R
q
( s)[ _ Cc
q
(1 [z[
d
)e
kj:j
.
and which depends smoothly on s.
In the estimates below, assumptions (5.2)(5.3) will be used. On one hand, by
(5.2) we have 1
1
estimates on , }
2
, and their rst derivatives; on the other, by
(5.3) we have 1
2
estimates on the higher-order derivatives, of the type |
(/)
|
1
2 _
C
/
(
/
c
/
)||
1
2, for l N.
We will also use notation like R
q
(.
0
), }
00
2
R
q
(.
0
), etc., to denote error
terms that are products of functions of s, like or }
0
2
, and the above R
q
s. Having
dened this notation, we can compute (and estimate) S
t
(

2,t
) term by term.
5.1 Estimate of A
1
From the expression of the Laplace-Beltrami operator (see Subsection 3.1) it
follows that
e
i
"
2
f
2
."s/
"
S
t
(

1,t
) S
t
(

[
1,t
)
= g
11
c
4
(}
0
2
)
2
[
1,t
i c
3
}
00
2

[
1,t
2i c
2
}
0
2
d
x

[
1,t
2i
/
g
1/
c
2
}
0
2
d
/

[
1,t

i
det g
d
g
1
det g
c
2
}
0
2

[
1,t
.
Using the expressions of n
i
. n
i
and the expansions of the metric coefcients in
Subsection 3.1 and multiplying the last equation by e
i(
Q
(
0
(tx))t
, we obtain
(5.17)
e
i
Q
f ."s/
"
A
1
= e
i
Q
f
0
."s/
"
e
i
"
2
f
2
."s/
"
S
t
(

1,t
) S
t
(

[
1,t
)
= e
i
Q
f ."s/
"
S
t
(

1,t
) e
i
Q
f
0
."s/
"
S
t
(

[
1,t
)
= A
1,0

A
1
:= A
1,0
A
1,i,e
A
1,i,o
A
1,i,e
A
1,i,o
A
1,1
.
where
(5.18)
A
1,0
= 2c
2
}
0
}
0
2
hU.
A
1,i,e
= c
3
}
0
2
2}
0
1
hU 4'H. )}
0
hU 2}
0
n
i,e
|.
A
1,i,o
= c
3
}
0
2
4'H. z)}
0
hU 2}
0
n
i,o
|.
A
1,i,e
= i c
3
}
00
2
hU 2i c
3
}
0
2
}
0
n
i,e
h
0
U hk
0
VU z|.
A
1,i,o
= 2i c
3
}
0
}
0
2
n
i,o
.
A
1,1
= (}
0
2
)
2
R
4
(.
0
) }
00
2
R
5
(.
0
) }
0
2
00
R
4
(.
0
)
}
0
2
R
4
(.
0
).
5.2 Estimate of A
2
Reasoning as for the previous estimate and collecting the terms of order c
3
and
higher in S
t
(

[
1,t
), we obtain
(5.19) e
i
Q
f ."s/
"
A
2
= A
2,0

A
2
:= A
2,i,e
A
2,i,o
A
2,i,e
A
2,i,o
A
2,1
.
where A
2,0
= 0 and where the remaining terms are given by
A
2,i,e
= c
3
00
J
e
( s).
A
2,i,o
= 2c
3
hk'H. )
00
}
d
}
U c
3
}
0
h}
0
1
}
z
}
00
}
U
2c
3
}
02
h'H. )
00
}
z
}
U.
A
2,i,e
= 2i c
3
}
0
h
},/
0
/
00
}
d
/
(z
}
U) 2i c
4
}
0
h'H. z)
000
}
z
}
U.
A
2,i,o
= i c
3
00
}
z
}
(}
00
hU }
0
h
0
U }
0
hk
0
VU z) i c
3
000
J
o
( s)
2i c
4
}
0
h'H. )
000
}
z
}
U.
A
2,1
= R
3
( s) (
0
)R
3
(.
0
.
00
) R
4
(.
0
.
00
).
where J
e
( s) and J
o
( s) are, respectively, an even real function and an odd real
function in the variables z, with smooth coefcients in s = cs, and satisfying the
decay property [J
e
( s)[ [J
o
( s)[ _ C(1 [z[
d
)e
kj:j
.
5.3 Choice of Q v and Estimate of A
3
We choose the function in such a way to annihilate (roughly) one of the main
terms in (5.17), namely 2c
2
}
0
}
0
2
hU(kz). Hence we dene so that it solves
(5.20) L
i
= 2}
0
}
0
2
hU(kz).
Reasoning as for the denition of n
i
(see Subsection 3.2), we can explicitly deter-
min as
= 2}
0
}
0
2
h
U(kz).
With this denition, using the above estimates on the metric coefcients and the
expressions of error terms, the linear terms involving 1 in S
t
(

2,t
) can be written
as
(5.21)
e
i
Q
f ."s/
"
A
3
=A
3,0

A
3
:=A
3,0
A
3,i,e
A
3,i,o
A
3,i,e
A
3,i,o
A
3,1
.
where
A
3,0
= c
2
L
i
. (5.22)
A
3,i,e
= 2c
3
}
0
h}
0
2
2(}
0
)
2
'H. ) 'V
1
V. )
](] 1)h
;2
U
;2
n
i,e

U
4c
3
(}
0
)
2
}
0
1
h}
0
2

U(kz) 2c
4
}
0
}
000
2
h
U.
A
3,i,o
= 2c
3
}
0
h}
0
2
2(}
0
)
2
'H. z) 'V
1
V. z)
](] 1)h
;2
U
;2
n
i,o

U
2c
3
}
0
}
0
2
hk
}
H
}
d
}

U(kz)
2c
4
}
0
hk}
0
2
00
}
d
}

U.
A
3,i,e
= 4i}
0
c
3
d
N x
(h}
0
}
0
2

U)
2i c
3
}
0
}
0
2
}
00
h (] 1)h
;1
U
;1
n
i,e

U.
A
3,i,o
= 2(] 1)i c
3
}
0
}
0
2
h
;1
U
;1
n
i,o

U
4i c
3
(}
0
)
2
}
0
2
hk
0
}
d
}

U.
A
3,1
= }
0
2
R
4
(.
0
.
00
)| }
00
2
R
4
(.
0
) R
6
(.
0
.
00
)|
c
4
}
000
2
R
1
(.
0
)| R
4
(.
0
)}
0
2
}
0
2
(1 c
2
}
0
2
) c}
00
2
|
R
5
(.
0
)}
00
2
}
0
2
R
6
(.
0
.
00
)(}
0
2
)
2
.
5.4 Choice of v
0
and Estimate of A
4
In order to make the approximate solution as accurate as possible, we add a
correction c
2
0
in such a way to compensate (most of) the terms c
2
(

1
i,e
i

1
i,o
);
see Subsection 3.3. We notice that these terms contain parts that are independent
of , which we denote by

1
0
i,e
and

1
0
i,o
, and parts that are quadratic in or
its derivatives

1
i,e
and

1
i,o
, respectively. Since we will take of order c, we
regard the latter ones as higher-order terms, and we add corrections to cancel

1
0
i,e
and

1
0
i,o
.
Specically, we dene
0
i,e
and
0
i,o
by
L
i
0
i,e
=
1
2
(}
0
)
2
hU(kz)
/,n
d
2
/n
g
11
z
n
z
/
2(}
0
)
2
'H. n
i,o
z) (5.23)
4(}
0
)
2
hU(kz)'H. z)
2
2}
0
d
x
n
i,e
}
00
n
i,e

h
00
U(kz) 2h
0
k
0
VU(kz) z hk
00
VU(kz) z
h(k
0
)
2
V
2
U(kz)z. z|
1
2
/,n
d
2
/n
g
t}
z
n
z
/
hk
2
d
2
t}
U(kz) kh'H. z)H
n
1
2
/
d
2
n/
g
11
z
/
|d
n
U(kz)
hk
/
d
2
/n
g
n}
z
/
d
}
U(kz)
/
H
/
d
/
n
i,o
hk'H. z)H
/
d
/
U(kz) 'V
1
V. n
i,o
z)
1
2
(] 1)h
;2
U(kz)
;2
n
2
i,e
1
2
](] 1)h
;2
U(kz)
;2
n
2
i,o
1
2
n,}
d
2
n}
Vz
n
z
}
hU(kz).
L
i
0
i,o
= 2
}
00
hU(kz) 2}
0
h
0
U(kz) 2}
0
hk
0
VU(kz) z
'H. z)
i
H
}
d
}
n
i,e
2(}
0
)
2
'H. n
i,e
z) 2}
0
d
x
n
i,o
}
00
n
i,o
}
0
hk
}
d
}
U(kz)
/,n
d
2
/n
g
1}
z
n
z
/
}
0
hU(kz)
n
d
2
1n
g
11
z
n
}
0
h
},/
d
2
/}
g
1}
z
/
U(kz)
1
2
}
0
h
/
d
2
1/
g
11
z
/
U(kz)
(] 1)h
;2
U(kz)
;2
n
i,o
n
i,e
'V
1
V. n
i,e
z).
(5.24)
We notice that the right-hand side of (5.23) is even in z, and hence orthogonal to
the kernel of L
i
. As a consequence, the equation is indeed solvable in
0
i,e
; see the
comments after (1.17). The same comment applies to (5.24), where the right-hand
side is odd in z. Furthermore, the right-hand sides decay at innity at most like
(1 [z[
d
)e
kj:j
for some integer J, so the same holds true for
0
i,e
and
0
i,o
.
In conclusion, after some computations we nd
e
i
Q
f ."s/
"
A
4
= A
4,0

A
4
:= A
4,0
A
4,i,e
A
4,i,o
A
4,i,e
A
4,i,o
A
4,1
.
where
(5.25)
A
4,0
= c
2
L
i
0
i,e
i c
2
L
i
0
i,o
.
A
4,i,e
= c
3
J
4,i,e
( s). A
4,i,o
= c
3
J
4,i,o
( s).
A
4,i,e
= c
3
J
4,i,e
( s). A
4,i,o
= c
3
J
4,i,o
( s).
(5.26)
A
4,1
= R
4
(.
0
) (
0
)(1 }
0
2
)R
3
(.
0
) }
00
2
R
5
(.
0
)
(}
0
2
)
2
R
6
( s)
00
R
4
(.
0
).
As for J
e
( s) and J
o
( s) in A
2
, the J
4
s depend only on V. ,, and M and are
bounded above by C(1 [z[
d
)e
kj:j
.
5.5 Estimate of A
5
The term involving
in S
t
(

2,t
) is given by
A
5
= A
5,0

A
5
:= A
5,0
A
5,i,e
A
5,i,o
A
5,i,e
A
5,i,o
A
5,1
(5.27)
where
A
5,0
= L
i
7
(kz) c
2
00
7
(kz) 2c
0
W
}
0
i L
i
W
i c
2
00
W
2i c
0
}
0
7
.
A
5,i,e
= c}
00
W
2c
2
d7
0
k
0
V7
(kz) z
2c}
0
dW
0
k
0
VW
(kz) z
(] 1)ch
;2
U
;2
n
i,e
W
.
A
5,i,o
= c
}
H
}
d
}
7
2c'H. z)(}
0
)
2
7
c
2
00
7
2c
0
}
0
W
|
c'V
1
V. z)7
](] 1)ch
;2
U
;2
n
i,o
7
.
A
5,i,e
= c}
00
7
2c
2
dW
0
k
0
VW
(kz) z
2c}
0
d7
0
k
0
V7
(kz) z
(] 1)ch
;2
U
;2
n
i,e
7
.
A
5,i,o
= c
}
H
}
d
}
W
2c'H. z)(}
0
)
2
W
c
2
00
W
2c
0
}
0
7
|
c'V
1
V. z)W
(] 1)ch
;2
U
;2
n
i,o
W
.
(5.28)
The error term A
5,1
= A
5,1
(. . }
2
) satises the following estimates:
[A
5,1
(. . }
2
)[ _ C(c
2
c
2
[}
0
2
[ c
3
[}
00
2
[ c
3
[
00
[)(1 [z[
d
)e
kj:j
([[ c[
0
[ c
2
[
00
[ c
3
[
000
[).
[A
5,1
(. . }
2
) A
5,1
(

.

.

}
2
)[
_ C
c[

[ c[
0

0
[ c
3
[
00

00
[
c
2
[}
0
2

}
0
2
[ c
3
[}
00
2

}
00
2
[
(1 [z[
d
)e
kj:j
([[ c[
0
[ c
2
[
00
[ c
3
[
000
[)
C(c
2
c
2
[}
0
2
[ c
3
[}
00
2
[ c
3
[
00
[)
(1 [z[
d
)e
kj:j
[

[ c[
0

0
[ c
2
[
00

00
[ c
3
[
000

000
[
.
By the form of the function (see (4.42), (4.43), and (5.5)), its Fourier modes are
mainly concentrated around indices of order
1
t
. As a consequence, 1
2
norms of
functions like c, c
2
00
, c
3
000
, etc., can be controlled with the 1
2
norm of ; see
also the comments before (4.58).
5.6 Estimate of A
6
First of all, we notice that we are taking
0
and }
0
2
in H
1
(0. 1|), and hence
they belong to 1
1
(0. 1|). As a consequence, since we have the bound
||
1
1
(0,1])
c|
0
|
1
1
(0,1])
c
2
|
00
|
1
1
(0,1])
c
3
|
000
|
1
1
(0,1])
_ Cc
2
(which follows from (5.7) and the above comments), we have the estimate
(5.29) [1[ [J[ [G[ _ Cc
2
(1 [z[
d
)e
kj:j
.
If we then choose sufciently small (also recall the expressions of n
i
and n
i
and
(5.9)), we deduce that
[

2,t

1,t
[ _ [

1,t
[ in

D
t
.
This estimate implies that A
6
admits a uniform quadratic Taylor expansion in
[

2,1

1,t
[ and is bounded by [

1,t
[
;2
[

2,t

1,t
[
2
.
Specically, we can write
(5.30) A
6
= A
6,0

A
6
:= A
6,0
A
6,i,e
A
6,i,o
A
6,i,e
A
6,i,o
A
6,1
.
where
A
6,0
= A
6,i,e
= A
6,i,o
= A
6,i,e
= A
6,i,o
= 0.
A
6,1
= 1
4
(}
0
2
. .
0
. ).
(5.31)
where 1
4
(}
0
2
. .
0
. ) is a quantity satisfying the estimates
[1
4
(}
0
2
. .
0
. )[ _
C
c
4
(c
2
||
1
1 c|
0
|
1
1)([[ c[
0
[)
(1 [z[
d
)e
kj:j
.
[1
4
(}
0
2
. .
0
. ) 1
4
(

}
0
2
.

.

0
.

)[
_ C(c
2
[[ c[
0
[ [

[ c[

0
[)(1 [z[
d
)e
kj:j

c[

[ c[
0

0
[ c
2
[}
0
2

}
0
2
[ [

[ c[
0

0
[
.
5.7 Final Estimate of S
"
.
Q
2;"
/
By (5.16), in the above notation we have
e
i
Q
f ."s/
"
S
t
(

2,t
) = c
2
(

1
i,o

1
i,e
) c
2
(

1
i,o,(
1

1
i,e,(
1
)
c
2
i(

1
i,e

1
i,o
) c
2
i(

1
i,e,(
1

1
i,o,(
1
)
iD1
A
i,0

6
iD1
A
i
.
Recalling the choices of ,
0
i,e
, and
0
i,o
in (5.20), (5.23), and (5.24) (and recalling
the notation for the 1s after (5.1)), we nally obtain the following result:
PROPOSITION 5.2 Suppose , }
2
, and satisfy (5.2), (5.3), and (5.7) for some
c
1
. c
2
. c
3
> 0. Let

} = } c}
1
c
2
}
2
, where } is given in (1.12) and }
1
in
(3.25). Let also n
i
= n
i,e
n
i,o
, with n
i,e
and n
i,o
given, respectively, in (3.13)
and (3.12), and n
i
= n
i,e
n
i,o
, where n
i,e
and n
i,o
are given in (3.9). Let

2,t
be dened in (5.8). Then, as c tends to 0, we have
(5.32)
e
i
Q
f
"
S
t
(

2,t
) = c
2
(

1
i,e

1
i,o

1
i,e,(
1

1
i,o,(
1
)
c
2
i(

1
i,e

1
i,o

1
i,e,(
1

1
i,o,(
1
)
L
i
7
(kz) c
2
00
7
(kz) 2c
0
W
}
0
i L
i
W
i c
2
00
W
2i c
0
}
0
7
e
i
Q
f
"
6
}D1
A
}
.
where the 1s are as in (5.1), where

1
i,e
and

1
i,o
are the terms quadratic in
and
0
within

1
i,e
and

1
i,o
, and where the latter error terms are given in (5.17),
(5.19), (5.21), (5.26), (5.27), and (5.30).
Remark 5.3. In some of the error terms listed above, we sometimes see derivatives
of order higher than 2 appearing on , }
2
, and . However, we are not only
assuming H
2
bounds on these functions, but also that they are linear combinations
of eigenfunctions corresponding to suitable eigenvalues. This fact then allows us
to derive bounds on higher-order norms; see the comments after (4.81) and (5.3)
and the last comments in the part concerning A
5
.
6 Proof of Theorem 1.1
In this section we prove our main theorem. First we solve the equation in the
H
t
components (see (4.87)), using a Lyapunov-Schmidt reduction. Then we turn
to the components in

1
and solve the bifurcation equation as well; this last step

crucially depends on the nondegeneracy assumption on , and an accurate choice
for the values of the parameter c.
6.1 Solvability in the Component of
N
H
"
In Proposition 4.4 we showed that problem(1.16) is reduced to nding a solution
of 1
t
() =

S
t
() in

D
t
(see (4.7), (4.23), and (4.24)) if we take 1
2
(cs) = V(cs).
Choosing

[
t
=

2,t
(the function constructed in the previous subsection) as an
approximation to the solution of Proposition 4.14, we have the following result
where, as usual, is sufciently small. We recall Proposition 4.4, formulas (4.85)
(4.88), and the denition of

1
after (4.86). Also, we denote by

t
the orthogonal
projection onto the set {e
i(
Q
( (tx)t)
:

1
].
PROPOSITION 6.1 Let

2,t
be as in Proposition 5.2. Then there exists `
,
depending on the parameters , }
2
, and , such that the following problem admits
a solution:
(6.1)
^
v
"
V(c.)
2,t
[
;1
(] 1)[

2,t
[
;3

2,t
m(

2,t
)
=

S
t
(
) e
i
Q
f ."s/
"
` .

H
t
. `
.
Furthermore, if m N and if

2,t
is an approximate solution corresponding to
different , }
2
, and , for a xed constant C independent of c and , for t =
1
2
and 0 < c
0
< c < 1 sufciently small, we have
|
0
,V
_
C
2
|

t
S
t
(

2,t
)|
1
2
(C
&;V
)
Cc
n
.
|` |
1
2
(C
&;V
)
_ C|S
t
(

2,t
)|
1
2
(C
&;V
)
.
(6.2)
|
0
,V
_
C
t
(S
t
(

2,t
) S
t
(

2,t
))
1
2
(C
&;V
)
. (6.3)
PROOF: The proof relies on Proposition 4.1, Proposition 4.14, and the contrac-
tion mapping theorem. By Proposition 4.14, the operator 1
t
(see (4.7)) is invert-
ible from (

H
t
. | |
0
,V
) into 1
2
(C
r
,V
), and the norm of the inverse is uniformly
bounded by C
2
. By this invertibility, (6.1) is satised if and only if

is a xed
point of the operator
`
J
t
: (

H
t
. | |
0
,V
) (

H
t
. | |
0
,V
) dened by
`
J
t
(
) =1
1
t

t
(

S
t
(
))|
:=1
1
t
S
t
(

2,t
) N
t
(n
t
c(
)) [

2,t
[
;1
c(
)
(] 1)[

2,t
[
;3
2,t
m(

2,t
c(
) )
.
We recall that, in the last formula, c(
) is given by Proposition 4.1, while N

t
is
dened in (4.8).
Our next goal is to show that
`
J
t
is a contraction on a metric ball (in the | |
0
,V
norm) of radius
C
2
|

t
S
t
(

2,t
)|
1
2
(C
&;V
)
Cc
n
for C large enough and m an
arbitrary integer. Setting for simplicity
`
G
t
(
) = N
t
(n
t
c(
)) [

2,t
[
;1
c(
)
(] 1)[

2,t
[
;3
2,t
m(

2,t
c(
) ).
by the above invertibility we clearly nd
(6.4)
|
`
J
t
(
)|
0
,V
_
C
t
S
t
(

2,t
)|
1
2
(C
&;V
)
|
`
G
t
(
)|
1
2
(C
&;V
)
.
|
`
J
t
(
1
)
`
J
t
(
2
)|
0
,V
_
C
2
|
`
G
t
(
1
)
`
G
t
(
2
)|
1
2
(C
&;V
)
.
We next evaluate |
`
G
t
(
)|
1
2
(C
&;V
)
and show that it is superlinear in |
|
1
2
(C
&
0
;V
)
up to negligible terms. We rst make the following claim:
Claim. In the notation of (4.14), letting k
1
( s) = (c
0
)
2
V( s) we have
|
|
C
1;1=2
k
1
_ C|
0
,V
for some C > 0. Assuming the claim true and choosing c
00
< (c
0
)
2
, we can
apply Proposition 4.1 with t =
1
2
, k
0
( s) = c
V( s), k
1
( s) = (c
0
)
2
V( s), and
k
2
( s) = c
00
V( s) to nd
|c(
)|
C
1=2
k
2
_ C
e
inf
k
2
Ck
0
K
t
|S
t
(

2,t
)|
C
1=2
k
0
e
inf
k
2
Ck
1
K
t
|
C
1;1=2
k
1
.
(6.5)
From the expressions for n
i
, n
i
, , and
0
and formula (5.29), we can deduce that
[

2,t
[ _ Ce
k
0
j:j
; moreover, from the estimates in the proof of Proposition 5.2
we also nd that |S
t
(

2,t
)|
1
2
(C
&;V
)
0 as c 0. By (4.20) (recall that > 0),
the latter bounds on

2,t
, the previous claim, and (6.5), if m is an arbitrary integer
and if c
00
is sufciently close to 1 after some elementary computations, we deduce
|
`
G
t
(
)|
1
2
(C
&;V
)
_ C
|
1Cc
1
2
(C
&;V
)
|
|
;
1
2
(C
&;V
)
c
n
1 |
|
1
2
(C
&;V
)
.
Similarly, if |
1
|
1
2
(C
&;V
)
and |
2
|
1
2
(C
&;V
)
are nite, we also nd
|
`
G
t
(
1
)
`
G
t
(
2
)|
1
2
(C
&;V
)
_
C
max
/D1,2
/
|
c^(;1)
1
2
(C
&;V
)
c
n
2
|
1
2
(C
&;V
)
.
where the symbol stands for the minimum. The last inequality and (6.4) show
that
`
J
t
is a contraction, and we obtain (6.2); (6.3) follows similarly.
To prove the claim, we note that according to our previous notation, the norm
| |
0
,V
is evaluated using the variables (s. z), where the zs are dened in (3.1).
If we want to estimate the | |
C
1;1=2
k
1
norm instead, we should use Lipschitz with
respect to s and ..
Given s
1
. s
2
R and .
1
. .
2
R
n1
, we want to consider the difference
V
(s
1
. .
1
) V
(s
2
. .
2
). Recalling (3.1), we can write that
d
x
(s
1
. .
1
) d
x
(s
2
. .
2
) = d
x
(s
1
. z
1
(cs
1
)) d
x
(s
2
. z
1
(cs
1
))
d
x
(s
2
. z
1
(cs
1
)) d
x
(s
2
. z
2
(cs
2
)).
By the denition of | |
0
,V
, d
x
H
1
(C
r
0
,V
) _ C
1
2
(C
r
0
,V
). This fact, the
smoothness of V( s), and ||
1
|
0
|
1
_ C(c
1
)c (which follows from (5.2))
imply that if (s
1
. .
1
). (s
2
. .
2
) T
1
(s. .), then
e
(
0
)
2
p
V(N x)j:j
[d
x
(s
1
. .
1
) d
x
(s
2
. .
2
)[ _
C(c
1
)|
0
,V
([s
1
s
2
[
1
2
[z
1
z
2
[
1
2
c[s
1
s
2
[).
A similar estimate holds for the derivatives of

with respect to ., so from (4.14)
we get the conclusion.
To apply Proposition 6.1, we establish explicit estimates on

t
S
t
(

2,t
) and
t
(S
t
(

2,t
) S
t
(

2,t
)). Specically, assuming from now on t =
1
2
, we have
the following result:
PROPOSITION 6.2 Assume , }
2
, ,

,

}
2
, and

satisfy conditions (5.2), (5.3),
and (5.7). Then, if

is dened as in Proposition 6.1, we have the estimates
c
2
|
(. . }
2
)|
0
,V
_ C(c
1
. c
2
. c
3
)c
3
. (6.6)
c
2
|
(. . }
2
)

(

.

.

}
2
)|
0
,V
_
C(c
1
. c
2
. c
3
)c
2
|

|
1
2 c
3
|}
2

}
2
|
1
2 c|

|
r
|.
(6.7)
where C(c
1
. c
2
. c
3
) is a positive constant depending on c
1
, c
2
, and c
3
but indepen-
dent of c and .
PROOF: We prove (6.6) only; (6.7) will follow from similar considerations. To
show (6.6) we use Proposition 6.1, so we are reduced to estimating
|

t
S
t
(

2,t
)|
1
2
(C
&;V
)
.
for which we can employ (5.32).
By our assumptions on , }
2
, and and by the estimates of the previous sub-
section, it is easy to see that
c
2
(

1
i,e

1
i,o

1
i,e,(
1

1
i,o,(
1
)
c
2
i(

1
i,e

1
i,o

1
i,e,(
1

1
i,o,(
1
)
1
2
(C
&;V
)
_
C(c
1
. c
2
. c
3
)c
3
c
.
e
i
Q
f
"
6
iD1
A
i
1
2
(C
&;V
)
_
C(c
1
. c
2
. c
3
)c
3
c
.
Recall that, by the above choices of
0
i,e
and
0
i,o
in (5.23) and (5.24), we have
corrected all the terms in the equation of order up to c
2
, so we are left with terms of
order c
3
and higher. The factor

c in the denominator arises from the fact that the
length of ,
t
is
1
t
; this gives a factor
1
t
when computing the 1
2
norm squared, and
we then need to take the square root. For the estimates in

A
6
, which also require
the 1
1
norm of , we can use the interpolation inequalities
||
1
1
(0,1])
_ C||
1
2
1
2
(0,1])
|
0
|
1
2
1
2
(0,1])
_ Cc
3
2
.
|
0
|
1
1
(0,1])
_ C|
0
|
1
2
1
2
(0,1])
|
00
|
1
2
1
2
(0,1])
_ Cc
1
2
.
It now remains to consider the other terms in the right-hand side of (5.32) in-
volving the functions 7
and W
. Let us call

1
1
t
the operator obtained from 1
1
t
(see (4.35)) by replacing the variables . with z and } with

} . Let us rst notice
that the terms under interest, with this notation, are nothing but

t

1
1
t
.
Let us now recall the expression of in (5.5) and
in (5.4); if
3,}
stands for
the functions in 1
3,
(see (4.49)) replacing . with z, we dene the function

2
t
}D
2
t
b
}

3,}
.
From the expression of
3,}
(see (4.45)), we nd that
|
|
,V
_
C

2
t
}D
2
t
b
2
}
c
2
(1 ;
2
)
1
2
. (6.8)
1
1
t
e
i
Q
f ."s/
"
=

1
1
t
(e
i
Q
f ."s/
"

)
O

2
t
}D
2
t
b
2
}
c
2
(1 ;
2
)
1
2
(in the | |
1
2
(C
&;V
)
norm).
(6.9)
Similarly to (4.68), recalling the asymptotic of v
}
(see (4.42) and the lines before)
we nd that
(6.10)

1
1
t
(e
i
Q
f ."s/
"

) = e
i
Q
f ."s/
"
2
t
}D
2
t
v
}
b
}

3,}
1
1
.
where
|1
1
|
1
2
(C
&;V
)
_
C

2
t
}D
2
t
b
2
}
c
2
(1 ;
2
)
1
2
_ C
c||
r
.
This implies the conclusion, by (5.7).
6.2 Projections onto
Q
K
In this section we estimate the projections of the equation onto the components
of

1
. We rst estimate their size and their Lipschitz dependence in the data ,
}
2
, and . Then we use the contraction mapping theorem to annihilate the function
` in Proposition 6.1, which implies the solvability of (1.16).
Projection onto
Q
K
1;
We want to evaluate the

1
1,
component of the function `
in (6.1). To do this
we consider a normal section to , that satises the rst relation in (5.3), and the
function
:= h(cs)
pC1
4
'(cs). V
:
U(kz))
i c'
0
(cs). z)
}
0
k
U(kz)
c
2
k
2
'
00
(cs). V(kz))
.
We then multiply both the left-hand side of (6.1) and

S
t
(
) (see (4.24)) by the

conjugate of e
i(
Q
( (tx))t
, integrate over

D
t
, and take the real part. When mul-
tiplying the left-hand side, we can integrate by parts and let the operator 1
t
act
on e
i(
Q
( (tx))t
. Using the arguments in the proofs of Proposition 4.9 (see in

particular (4.66) and (4.67)) and of Proposition 6.2, we nd that
1
t
e
i
Q
f ."s/
"
= e
i
Q
f ."s/
"

1(
).
where

1
1,
and where
|1(
)|
1
2
(C
&;V
)
_ C(c
3
)|
|
1
2
(C
&;V
)
_
C
c
(c
3
)||
1
2
(0,1])
.
Therefore, since

is orthogonal to

1
, from (6.6) we deduce that

(6.11)
Q
T
"
e
i
Q
f ."s/
"
1
t
JV
Q v
"
_
C
c
(c
3
)||
1
2
(0,1])
|
|
1
2
(C
&;V
)
_ C(c
1
. c
2
. c
3
)c
2
||
1
2
(0,1])
.
We next have to consider

S
t
(
), whose main term is S

t
(

2,t
). For this we use
formula (5.32). Here we have three kinds of terms: the

1s, those involving 7
and W
(which coincide with A

5,0
, with our notation in (5.28)), and the

As.
For the

1s, since
is odd in z, the products with the even terms will vanish.

The products of the odd terms (notice that the two phases cancel and we use the
change of variables s cs) instead give us
c
2
m
Q
T
"
(

1
i,o

1
i,o,(
1
)
JV
Q v
"
c
2
m
Q
T
"
i(

1
i,o

1
i,o,(
1
)
JV
Q v
"
= c
] 1
20
C
0
1
0
'J(). ) J s

1
0
.
where C
0
=

R
n1
U(.)
2
J. and [

1
0
[ _ Cc||
1
2
(0,1])
. To explain why this
estimate holds, we notice rst that
;1
20
C
0
'J(). ) is exactly the rst term of
multiplied by

1
i,o

1
i,o,(
1
, as shown in Subsection 3.3 (the factor h
(;C1)4
in
(4.39) is needed precisely to cancel the factor
1
hk
in (3.26)). The remaining terms
in the last equation are given either by products of the imaginary part of
and the
imaginary

1s or that of

1
i,o

1
i,o,(
1
and the last term in
. In the latter case,

for example (see the comments after (4.81)), we obtain a quantity bounded by
Cc
2
1t
0
([[ [
0
[ [
00
[)c
2
00
Js _ Cc
2
2
||
1
2
(0,1])
.
The last inequality follows from (5.2) and the fact that satises the rst condition
in (5.3). On the other hand, the terms involving
0
once integrated will be bounded
by Cc
2
||
1
2
(0,1])
, still by (5.3).
Concerning A
5,0
, we next claim that for any m N we have
(6.12)
Q
T
"
A
5,0
JV
Q v
"
_ Cc
n
||
1
2
(0,1])
as c 0.
To see this, notice that satises (5.3), while A
5,0
arises from functions involving
(in particular , see (5.5)); since ; ranges between

2
c and
2
c, the main
modes of are much higher than those for . Hence, using Fourier cancellation
as in Lemma 4.10, we can deduce (6.12). It is also easy to see that
(6.13)
Q
T
"
}D1
A
}
JV
Q v
"
_ C(c
1
. c
2
. c
3
)c
2
||
1
2
(0,1])
.
Finally, it remains to consider the product of
and the last three terms in

(4.24). Indeed, since these are either superlinear in

(see (4.20)) or contain c(
)
(see (4.15)), they are of lower order compared to (6.11).
Using (6.11)(6.13) and the above arguments, we nally obtain that, if ` is as in
Proposition 6.1, then
Q
T
"
`
JV
Q v
"
= c
] 1
20
C
0
1
0
'J(). )J s 1
1
.
[1
1
[ _ C(c
1
. c
2
. c
3
)c
2
||
1
2
(0,1])
.
Similarly, using the estimates in Section 5, we nd that if

` corresponds to the
triple (
,

}
2
,

), then
(6.14)
Q
T
"
( `

` )
JV
Q v
"
= c
] 1
20
C
0
1
0
'J(

). )J s

1
1
.
where

1
1
satises
[

1
1
[ _ C(c
1
. c
2
. c
3
)
c|

|
1
2
(0,1])
c
2
|}
2

}
2
|
1
2
(0,1])
|

|
r
||
1
2.
(6.15)
Projection onto
Q
K
2;
For this projection we will be sketchier since most of the arguments of the pre-
vious projection can be applied. If }
2
satises the second condition in (5.3), we
consider the function
(
2
= h(cs)
1
2
i }
2
(cs)U(kz) 2c
}
0
}
0
2
(cs)
k
U(kz) i c
2
}
00
2
(cs)
k
2
W(kz)
.
As for the previous case, the main contribution to the projection is given by the
product of the rst term in
(
2
and the imaginary parts of S
t
(

2,t
) listed in (5.32),
which are even in z.
We denote by

1
i,e,(
2
the sum of all imaginary even terms of order c
3
appearing
in the equation, namely A
1,i,e
, A
3,i,e
, and A
4,i,e
= J
4,i,e
( s) (see (5.18), (5.22),
and (5.25))
(6.16)
1
i,e,(
2
=2h
0
}
0
2
U 2h}
0
2
k
0
VU z 2}
0
}
0
2
n
i,e
}
00
2
hU
4}
0
d
x
(h}
0
}
0
2

U) 2}
00
h}
0
}
0
2

U
2(] 1)h
;1
[U[
;2
}
0
}
0
2

Un
i,e
J
4,i,e
( s)
:=

1
i,e,(
2
J
4,i,e
( s).
Notice that

1
i,e,(
2
coincides with the function

1
i,e,(
1
in (5.1) (see Subsection 3.3
for the precise expression) if we replace }
1
with }
2
. Therefore, from estimates
similar to the previous ones (which mainly use the computations in subsection 4.1
in [36]), we nd
Q
T
"
`
(
2
JV
Q v
"
= c
2
C
0
1
0
T(}
2
)}
2
J s
c
2
1
0

R
n1
J
4,i,e
U(k( s))
}
2
J s 1
2
.
(6.17)
where C
0
=

R
n1
U(.)
2
J.,
(6.18) T(}
2
) = d
N x
h
2
}
0
2
(] 1)k
nC1
(] 1)h
;1
2oA
2
h
2c
|
.
and 1
2
satises
(6.19) [1
2
[ _ C(c
1
. c
2
. c
3
)c
2
|}
2
|
1
2
(0,1])
.
Moreover, if

` corresponds to the triple (
.

}
2
.

), then
(6.20)
Q
T
"
( `

` )
(
2
JV
Q v
"
= c
2
C
0
1
0
T(}
2

}
2
)}
2
J s

1
2
.
with
[

1
2
[ _ C
c
2
|}
2

}
2
|
1
2
(0,1])
c|

|
1
2
(0,1])
|

|
r
|}
2
|
1
2
(0,1])
(6.21)
with C = C(c
1
. c
2
. c
3
).
Projection onto
Q
K
3;
To compute the last components of the projection, we recall our notation in
Subsection 4.3 and dene
(cs) =
2
t
}D
2
t
b
}
}
(cs).
2
t
}D
2
t
b
}

3,}
.
As for the previous cases, the main contribution to the projection here still comes
from S
t
(

2,t
). In particular, following the arguments for

1
1,
, when testing on
, by Fourier cancellation and parity the major terms are indeed A

5,0
, A
5,i,e
, and
A
5,i,e
. With straightforward computations, we nd that
(6.22)
Q
T
"
`
JV
Q v
"
=
1
c
1
0
(. . . )J s

1
3
.
where
(. . . ) = O
4,
c
2
00
O
1,
2c
0
}
0
O
3,
O
5,
c
2
00
O
2,
c}
00
( )O
3,
2c
0
}
0
O
3,
2c
2
0
(
0
O
6,

0
O
T,
) 2c
2
k
0
(
0
O
10,

0
O
11,
)
2c}
0
k
0
(O
12,
O
13,
) 2c}
0
0
(O
S,
O
9,
)
(] 1)ch
;2
( )O
14,
.
O
4,
( s) =
R
n1
7
(N x)
L
i
7
(N x)
. O
5,
( s) =
R
n1
W
(N x)
L
i
W
(N x)
.
O
6,
( s) =
R
n1
7
(N x)
d7
(N x)
d
. O
T,
( s) =
R
n1
W
(N x)
dW
(N x)
d
.
O
S,
( s) =
R
n1
7
(N x)
dW
(N x)
d
. O
9,
( s) =
R
n1
W
(N x)
d7
(N x)
d
.
O
10,
( s) =
R
n1
7
(N x)
V
:
7
(N x)
z. O
11,
( s) =
R
n1
W
(N x)
V
:
W
(N x)
z.
O
12,
( s) =
R
n1
7
(N x)
V
:
W
(N x)
z. O
13,
( s) =
R
n1
W
(N x)
V
:
7
(N x)
z.
and
O
14,
( s) =
R
n1
U(kz)
;2
n
i,e
W
(N x)
7
(N x)
.
[

1
3
[ _ C(c
1
. c
2
. c
3
)c
2
||
1
2
(0,1])
. (6.23)
After some manipulation using the fact that (7
. W
) solves (4.29) with n
= 0,
the normalization

R
n1
(7
2
W
2
) = 1, and some integration by parts in z, we

nd that
(6.24)
1
c
1
0
(. . . )J s =
1
c
1
0
0
(. . . )J s
1
0
1
(. . . )J s.
where
0
(. . . ) = O
1,
(c
2
0
k
2
2
) O
2,
(c
2
0

2
k
2
)
2}
0
O
3,
(c
0
c
0
k k)
(6.25)
and
1
= ( )g( s)
with
g( s) = }
00
O
3,
2}
0
k
0
O
13,
2}
0
0
O
9,
(] 1)h
;2
O
14,
|.
Now we notice that, by (4.43), we have
=
c
k

2
t
},/D
2
t
b
}
b
/
(
0
}
/

}
0
/
)
J
1
2
t
},/D
2
t
b
}
b
/
(v
}
0
}
v
/
0
/
)
.
where
J
1
=
O
1,
k
2
2
2}
0
kO
3,
.
Integrating by parts in s and using (4.42), we nd that
(6.26)
1
0
1
(. . . )J s =
c
1
0
g
k
0
( s)J s O(
2
)||
1
2
(0,1])
||
1
2
(0,1])
.
Finally, combining (6.22), (6.23), (6.24), and (6.26), we deduce
(6.27)
Q
T
"
`
JV
Q v
"
=
1
c
1
0
0
(. . . )J s 1
3
.
where
(6.28) [1
3
[ _ C(c
1
. c
2
. c
3
)(c
2
(c
2
)||
1
2
(0,1])
)||
1
2
(0,1])
.
Analogously we obtain
(6.29)
Q
T
"
( `

` )
JV
Q v
"
=
1
c
1
0
0
(

.

. . )

1
3
.
where

1
3
satises
[

1
3
[ _ C
c|

|
1
2
(0,1])
c
2
|}
2

}
2
|
1
2
(0,1])
|

|
r
||
1
2
(0,1])
(6.30)
with C = C(c
1
. c
2
. c
3
).
Remark 6.3. Let us consider the eigenvalue problem in (. )
1
0
0
(. . . ) = v
1
0
(O
1,
O
2,
) for all (. )
where O
1
and O
2
are dened in (4.41). Then the eigenvalue equation is the fol-
lowing:
(6.31)
c
2
(O
1;
0
)
0
O
1;
k
2
2
2}
0
O
3;
O
1;
(c
0
k) = v.
c
2
(O
2;
0
)
0
O
2;
k
2
2
2}
0
O
3;
O
2;
(c
0
k) = v.
By (4.44), the couple of functions (
}
.
}
) constructed in Subsection 4.3 represents
a family of approximate eigenfunctions corresponding to v = v
}
.
6.3 The Contraction Argument
The usual procedure in performing a xed-point argument is to apply to the
equation an invertible linear operator rst. In the expansions in the last subsection,
we showed that the main terms in the projections onto

1
are the operators J, T ,

and
0
(the last is identied by duality with the associated quadratic form); see
(6.2), (6.17), and (6.27). By our nondegeneracy assumption on ,, J is invertible
and the same holds also for T , since it is coercive (and in divergence form). It
remains then to invert
0
, which is the content of the next result. Before stating it
we introduce some notation. Using the symbology of Subsection 4.3 we dene the
spaces
X
1,
= span
c
}
: ; = 0. . . . .
. X
2,
= span
o
}
: ; = 0. . . . .
.
X
3,
= span
}
: ; =
2
c
. . . . .
2
c
.
with X
1,
and X
2,
endowed with the H
2
norm on 0. 1|, and X
3,
with the | |
r
norm.
We also call Y
1,
, Y
2,
, and Y
3,
the same spaces of functions, but endowed
with weighted 1
2
norms: by the normalization after (4.38), it is natural to put
the weights h
0
and h
c
on Y
1,
and Y
2,
, respectively. Concerning Y
3,
, by
Remark 6.3, we will endow it with the product (. )
Y
3;
=

1
0
(O
1,

O
2,
)J s where, as above, is related to by (4.43) and (5.5). Notice that
by (4.38) J and T are exactly diagonal from X
1,
to Y
1,
and from X
2,
to Y
2,
,
respectively, while
0
is nearly diagonal (see also (4.44)).
LEMMA 6.4 Letting
Y
3;
denote the orthogonal projection onto Y
3,
, there exists
a sequence c
k
0 such that
0
is invertible from X
3,
into Y
3,
and such that its
inverse satises |(
Y
3;
0
)
1
| _
C
t
k
for some xed constant C.
PROOF: First of all, we show that there exists c
k
0 such that
Y
3;
0
cannot have eigenvalues in Y
3,
smaller in absolute value than C
1
c
k
; after this,
we estimate the (stronger) X
3,
norm of its inverse.
To prove the claim, we apply Katos theorem (see [29, p. 445]), which allows
us to compute the derivative of an eigenvalue v(c) of
Y
3;
0
with respect to c.
The (possibly multiple) value of this derivative is given by the eigenvalues of
Y
3;
d
t
0
restricted to the v(c)-eigenspace of
Y
3;
0
.
Suppose that satises the eigenvalue equation
Y
3;
0
= v, which is
equivalent to
(6.32)
1
0
0
(. . . ) = v
1
0
(O
1,
O
2,
)
for all (. )with Y
3,
.
Looking at the powers of c in
0
(see (6.25)), we write
0
=
0,0
c
0,1

c
2
0,2
; notice that
0,0
is negative denite and
0,2
positive denite. We also
point out that, since }
0
satises (1.12), for }(cs)c to be 1c-periodic, when we
vary c, A also needs to be adjusted. Specically, since the total variation of phase
in (1.10) is
A
1t
0
h(cs)
c
Js =
A
c
1
0
h( s)J s = const.
when differentiating with respect to c we nd that
JA
Jt
=
A
t
. Hence, applying
Katos theorem, we nd
(6.33)
dv
dc

min
1
,
2
0
(
1
.
2
). max
1
,
2
0
(
1
.
2
)
.
where
(
1
.
2
) =
1
0
(
0,1
2c
0,2
)(
1
.
1
.
2
.
2
)
1
0
(O
1,
2
O
2,
2
)
1
c
1
0
(2}
0
O
3,
(c
0
1
2
c
0
1
2
k
1
2
k
1
2
)
1
0
(O
1,
2
O
2,
2
)
.
and where (
1
.
1
). (
2
.
2
) are functions satisfying (6.32). By using this, we can
write (
1
.
2
) as
1
t
1
0
(
0

0,0
)(
1
.
1
.
2
.
2
) c
1
0

0,2
(
1
.
1
.
2
.
2
)
1
0
(O
1,
2
O
2,
2
)
1
c
1
0
(2}
0
O
3,
(c
0
1
2
c
0
1
2
k
1
2
k
1
2
)
1
0
(O
1,
2
O
2,
2
)
=
v
c
1
0
(O
1,
0
1
0
2
O
2,
0
1
0
2
)
1
0
(O
1,
2
O
2,
2
)
1
c
1
0
k
2
2
(O
1,
2
O
2,
2
) 2}
0
kO
3,
(
1
2

1
2
)|
1
0
(O
1,
2
O
2,
2
)
1
c
1
0
(2}
0
O
3,
(c
0
1
2
c
0
1
2
k
1
2
k
1
2
)
1
0
(O
1,
2
O
2,
2
)
.
Applying (4.42), (4.43), and O
1,
O
2,
= 1 (see (4.41) and the lines after
(4.30)), the last expression simplies to
v
c
1
c
1
0
(2
2
k
2
4}
0
kO
3,
)
1
1
0

1
2
O(
2
)
1
c
.
Since the numerator is symmetric in
1
and
2
, the inmum of the above ratio is
realized by some
0
, so by (6.33) and the latter formula we nd
(6.34)
dv
dc
_
v
c
1
c
1
0
(2
2
k
2
4}
0
kO
3,
)
2
0
1
0

2
0
O(
2
)
1
c
_
1
c
v inf
0,1]
(2
2
k
2
4}
0
kO
3,
) C
2
|.
Notice that for v and sufciently small, the coefcient of
1
t
in the above for-
mula is positive and uniformly bounded away from 0. From (4.44) and the asymp-
totics in (4.42) (which follows from Weyls formula), we can show that
Y
3;
0
has a number of negative eigenvalues of order
2
c. This fact and (6.34) yield
the desired claim, which can be obtained as in [41, prop. 4.5]; since the argument
is quite similar, we omit the details. The above claim provides invertibility of
Y
3;
0
in Y
3,
and gives
(6.35) |(
Y
3;
0
)
1
|
Y
3;
_
C
c
||
Y
3;
for any Y
3,
.
We next want to estimate the X
3,
norm of (
Y
3;
0
)
1
. Let
=
2
t
}D
2
t
b
}
}
and suppose

=

2
t
}D
2
t
b
}
}
is such that
Y
3;
= in the sense that
1
0
0
(
. . ) =
1
0
(O
1,
O
2,
) for all (. ) with Y
3,
.
If =

2
t
}D
2
t
b
}
}
, then by (4.44), we nd by integrating
2
t
}D
2
t
v
}

b
}
b
}
O

2
t
}D
2
t
(v
2
}
c)
2
b
2
}
1
2

2
t
/D
2
t
b
2
/
1
2
_
C

2
t
}D
2
t
b
2
}
1
2

2
t
/D
2
t
b
2
/
1
2
.
Choosing b
}
=

b
}
for ; > 0 and b
}
=
b
}
for ; < 0, from the asymptotics of
v
}
in (4.42), we obtain for

C
1
> 0 sufciently large that
c
Q
C
1
}
2
t
[; [
b
2
}
_ C
2
t
}D
2
t
b
2
}
_ C|
|
2
Y
3;
.
By (6.35) we have |
|
Y
3;
_
C
t
||
Y
3;
, so recalling (5.6), we get
|
|
2
A
3;
:= |
|
2
r
_ C|
|
2
Y
3;
_
C
2
c
2
||
2
Y
3;
.
which yields the conclusion.
PROOF OF THEOREM 1.1: Let us introduce the operators
G
/
: X
1,
X
2,
X
3,
Y
/,
. l = 1. 2. 3.
dened by duality as
(G
1
(. }
2
. ). )
Y
1;
=
Q
T
"
`
JV
Q v
"
.
(G
2
(. }
2
. ). }
2
)
Y
2;
=
Q
T
"
`
(
2
JV
Q v
"
.
(G
3
(. }
2
. ). )
Y
3;
=
Q
T
"
`
JV
Q v
"
.
where ` = ` (. }
2
. ) is the function appearing in Proposition 6.1.
By Proposition 4.4, equation (1.16) (or (NLS
t
)) is solved if and only if ` = 0.
In the above notation, this is equivalent to nding (. }
2
. ) such that G
/
(. }
2
. )
= 0 for every l = 1. 2. 3. If c
k
is the sequence given in Lemma 6.4, then
0
is
invertible, and the condition ` = 0 is equivalent to the system (we set c = c
k
)
(6.36)
= G
1
(. }
2
. ) :=
1
t
J
1
G
1
(. }
2
. ) c
J()|.
}
2

`
}
2
= G
2
(. }
2
. )
:=
1
t
2
T
1
G
2
(. }
2
. ) c
2

T }
2
c
2
R
n1
J
4,i,e
U(k( s)z)Jz
.
= G
3
(. }
2
. ) := c(
Y
3;
0
)
1
G
3
(. }
2
. )
1
t
Y
3;
.
where
J =
] 1
20
C
0
J.

T = C
0
T.
with (C
0
=

R
n1
U(.)
2
J.) and where
`
}
2
=

T
1

R
n1
J
4,i,e
U(k( s)z)Jz
.
By (6.2)(6.21), (6.27)(6.29), and (6.30) we nd
|G
1
(0. 0. 0)|
A
1;
_ Cc. |G
2
(0. 0. 0)|
A
2;
_ C. |G
3
(0. 0. 0)|
A
3;
_ Cc
2
:
moreover, if , }
2
, and satisfy the bounds (5.2) and (5.7), then
|G
1
(. }
2
. ) G
1
(
.

}
2
.

)|
A
1;
_
C(c
1
. c
2
. c
3
)
|

|
A
1;
c|}
2

}
2
|
A
2;

c
|

|
A
3;
.
|G
2
(. }
2
. ) G
2
(
.

}
2
.

)|
A
2;
_
C(c
1
. c
2
. c
3
)
c
|

|
A
1;
|}
2

}
2
|
A
2;

c
2
|

|
A
3;
.
|G
3
(. }
2
. ) G
3
(
.

}
2
.

)|
A
3;
_
C(c
1
. c
2
. c
3
)
c|

|
A
1;
c
2
|}
2

}
2
|
A
2;
|

|
A
3;
.
We now consider the scaled norms
c| |
y
A
1;
= | |
A
1;
.
1
2
| |
y
A
2;
= | |
A
2;
. c
2
| |
y
A
3;
= | |
A
3;
.
With this new notation the last formulas become
(6.37)
|G
1
(0. 0. 0)|
y
A
1;
_ C. |G
2
(0. 0. 0)|
y
A
2;
_ C
1
2
.
|G
3
(0. 0. 0)|
y
A
3;
_ C.
|G
1
(. }
2
. ) G
1
(
.

}
2
.

)|
y
A
1;
_
C(c
1
. c
2
. c
3
)
|

|
y
A
1;
1
2
|}
2

}
2
|
y
A
2;
|

|
y
A
3;
.
|G
2
(. }
2
. ) G
2
(
.

}
2
.

)|
y
A
2;
_
C(c
1
. c
2
. c
3
)
1
2
|

|
y
A
1;
|}
2

}
2
|
y
A
2;
1
2
|

|
y
A
3;
.
|G
3
(. }
2
. ) G
3
(
.

}
2
.

)|
y
A
3;
_
C(c
1
. c
2
. c
3
)
|

|
y
A
1;
3
2
|}
2

}
2
|
y
A
2;
|

|
y
A
3;
.
If C is the constant appearing in (6.37), from the last four formulas we deduce that
if is sufciently small then (G
1
. G
2
. G
3
) has a xed point in
{| |
y
A
1;
_ 2C] {|
`
}
2
|
y
A
2;
_ 2C
1
2
] {| |
y
A
3;
_ 2C].
This, by the comments before (6.36), leads to a solution of (NLS
t
) with the desired
asymptotics.
Acknowledgment. This work started in the fall of 2004, when the second author
visited Departamento de Matemtica in Universidade Estadual de Campinas. He is
very grateful to the institution for the kind hospitality. F.M. and A.M. are supported
by Ministero dellUniversit a e della Ricerca within the PRIN 2006 Variational
methods and nonlinear differential equations. F.M. is grateful to SISSA for its
kind hospitality. M.M. was supported by FAEP-UNICAMP, FAPESP, and CNPq.
Bibliography
[1] Ambrosetti, A.; Badiale, M.; Cingolani, S. Semiclassical states of nonlinear Schrdinger equa-
tions. Arch. Rational Mech. Anal. 140 (1997), no. 3, 285300.
[2] Ambrosetti, A.; Felli, V.; Malchiodi, A. Ground states of nonlinear Schrdinger equations with
potentials vanishing at innity. J. Eur. Math. Soc. (JEMS) 7 (2005), no. 1, 117-144.
[3] Ambrosetti, A.; Malchiodi, A. Perturbation methods and semilinear elliptic problems on R
n
.
Progress in Mathematics, 240. Birkhuser, Basel, 2006.
[4] Ambrosetti, A.; Malchiodi, A.; Ni, W.-M. Singularly perturbed elliptic equations with symme-
try: existence of solutions concentrating on spheres. I. Comm. Math. Phys. 235 (2003), no. 3,
427466.
[5] Ambrosetti, A.; Malchiodi, A.; Ni, W.-M. Singularly perturbed elliptic equations with symme-
try: existence of solutions concentrating on spheres. II. Indiana Univ. Math. J. 53 (2004), no. 2,
297329.
[6] Ambrosetti, A.; Malchiodi, A.; Ruiz, D. Bound states of nonlinear Schrdinger equations with
potentials vanishing at innity. J. Anal. Math. 98 (2006), 317348.
[7] Ambrosetti, A.; Malchiodi, A.; Secchi, S. Multiplicity results for some nonlinear Schrdinger
equations with potentials. Arch. Ration. Mech. Anal. 159 (2001), no. 3, 253271.
[8] Arioli, G.; Szulkin, A. A semilinear Schrdinger equation in the presence of a magnetic eld.
Arch. Ration. Mech. Anal. 170 (2003), no. 4, 277295.
[9] Badiale, M.; DAprile, T. Concentration around a sphere for a singularly perturbed Schrdinger
equation. Nonlinear Anal. 49 (2002), no. 7, Ser. A: Theory Methods, 947985.
[10] Bartsch, T.; Peng, S. Semiclassical symmetric Schrdinger equations: existence of solutions
concentrating simultaneously on several spheres. Z. Angew. Math. Phys. 58 (2007), no. 5, 778
804.
[11] Benci, V.; DAprile, T. The semiclassical limit of the nonlinear Schrdinger equation in a radial
potential. J. Differential Equations 184 (2002), no. 1, 109138.
[12] Berestycki, H.; Lions, P.-L. Nonlinear scalar eld equations. I. Existence of a ground state.
Arch. Rational Mech. Anal. 82 (1983), no. 4, 313345.
[13] Byeon, J.; Wang, Z.-Q. Standing waves with a critical frequency for nonlinear Schrdinger
equations. Arch. Ration. Mech. Anal. 165 (2002), no. 4, 295316.
[14] Dancer, E. N.; Yan, S. Multipeak solutions for a singularly perturbed Neumann problem. Pacic
J. Math. 189 (1999), no. 2, 241262.
[15] Dancer, E. N.; Yan, S. A new type of concentration solutions for a singularly perturbed elliptic
problem. Trans. Amer. Math. Soc. 359 (2007), no. 4, 17651790 (electronic).
[16] DAprile, T. On a class of solutions with non-vanishing angular momentum for nonlinear
Schrdinger equations. Differential Integral Equations 16 (2003), no. 3, 349384.
[17] del Pino, M.; Felmer, P. L. Semi-classical states for nonlinear Schrdinger equations. J. Funct.
Anal. 149 (1997), no. 1, 245265.
[18] del Pino, M.; Felmer, P. L.; Wei, J. On the role of mean curvature in some singularly perturbed
Neumann problems. SIAM J. Math. Anal. 31 (1999), no. 1, 6379 (electronic).
[19] del Pino, M.; Kowalczyk, M.; Wei, J.-C. Concentration on curves for nonlinear Schrdinger
equations. Comm. Pure Appl. Math. 60 (2007), no. 1, 113146.
[20] do Carmo, M. Riemannian geometry. Mathematics: Theory and Applications. Birkhuser,
Boston, 1992.
[21] Floer, A.; Weinstein, A. Nonspreading wave packets for the cubic Schrdinger equation with a
bounded potential. J. Funct. Anal. 69 (1986), no. 3, 397408.
[22] Gidas B.; Ni, W. M.; Nirenberg, L. Symmetry of positive solutions of nonlinear elliptic equa-
tions in 1
n
. Mathematical analysis and applications, part A, 369402. Advances in Mathemat-
ics Supplementary Studies, 7a. Academic, New YorkLondon, 1981.
[23] Grossi, M. Some results on a class of nonlinear Schrdinger equations. Math. Z. 235 (2000),
no. 4, 687705.
[24] Grossi, M.; Pistoia, A.; Wei, J. Existence of multipeak solutions for a semilinear Neumann prob-
lem via nonsmooth critical point theory. Calc. Var. Partial Differential Equations 11 (2000),
no. 2, 143175.
[25] Gui, C. Multipeak solutions for a semilinear Neumann problem. Duke Math. J. 84 (1996), no. 3,
739769.
[26] Gui, C.; Wei, J. On multiple mixed interior and boundary peak solutions for some singularly
perturbed Neumann problems. Canad. J. Math. 52 (2000), no. 3, 522538.
[27] Gui, C.; Wei, J.; Winter, M. Multiple boundary peak solutions for some singularly perturbed
Neumann problems. Ann. Inst. H. Poincar Anal. Non Linaire 17 (2000), no. 1, 4782.
[28] Jeanjean, L.; Tanaka, K. A positive solution for a nonlinear Schrdinger equation on R
1
. Indi-
ana Univ. Math. J. 54 (2005), no. 2, 443464.
[29] Kato, T. Perturbation theory for linear operators. 2nd ed. Grundlehren der Mathematischen
Wissenschaften, 132. Springer, BerlinNew York, 1976.
[30] Kwong, M. K. Uniqueness of positive solutions of û u u
;
= 0 in R
n
. Arch. Rational
Mech. Anal. 105 (1989), no. 3, 243266.
[31] Li, Y. Y. On a singularly perturbed equation with Neumann boundary condition. Comm. Partial
Differential Equations. 23 (1998), no. 3-4, 487545.
[32] Li, Y.; Nirenberg, L. The Dirichlet problem for singularly perturbed elliptic equations. Comm.
Pure Appl. Math. 51 (1998), no. 11-12, 14451490.
[33] Lin, C.-S.; Ni, W.-M.; Takagi, I. Large amplitude stationary solutions to a chemotaxis system.
J. Differential Equations 72 (1988), no. 1, 127.
[34] Mahmoudi, F.; Malchiodi, A. Concentration on minimal submanifolds for a singularly per-
turbed Neumann problem. Adv. Math. 209 (2007), no. 2, 460525.
[35] Mahmoudi, F.; Malchiodi, A. Solutions to the nonlinear Schrdinger equation carrying momen-
tum along a curve. II. Proof of the existence result. Preprint. arXiv: 0708.0104v1, 2007.
[36] Mahmoudi, F.; Malchiodi, A.; Montenegro, M. Solutions to the nonlinear Schrdinger equation
carrying momentum along a curve. I. Study of the limit set and approximate solutions. Preprint.
arXiv: 0708.0125v1, 2007.
[37] Mahmoudi, F.; Malchiodi, A.; Montenegro, M. Solutions to the nonlinear Schrdinger equation
carrying momentum along a curve. C. R. Math. Acad. Sci. Paris 346 (2008), no. 1-2, 3338.
[38] Mahmoudi, F.; Mazzeo, R.; Pacard, F. Constant mean curvature hypersurfaces condensing along
a submanifold. Geom. Funct. Anal. 16 (2006), no. 4, 924958.
[39] Malchiodi, A. Concentration at curves for a singularly perturbed Neumann problem in three-
dimensional domains. Geom. Funct. Anal. 15 (2005), no. 6, 11621222.
[40] Malchiodi, A.; Montenegro, M. Boundary concentration phenomena for a singularly perturbed
elliptic problem. Comm. Pure Appl. Math. 55 (2002), no. 12, 15071568.
[41] Malchiodi, A.; Montenegro, M. Multidimensional boundary-layers for a singularly perturbed
Neumann problem. Duke Math. J. 124 (2004), no. 1, 105143.
[42] Mazzeo, R.; Pacard, F. Foliations by constant mean curvature tubes. Comm. Anal. Geom. 13
(2005), no. 4, 633670.
[43] Molle, R.; Passaseo, D. Concentration phenomena for solutions of superlinear elliptic problems.
Ann. Inst. H. Poincar Anal. Non Linaire 23 (2006), no. 1, 6384.
[44] Ni, W.-M. Diffusion, cross-diffusion, and their spike-layer steady states. Notices Amer. Math.
Soc. 45 (1998), no. 1, 9-18.
[45] Ni, W.-M.; Takagi, I. On the shape of least-energy solutions to a semilinear Neumann problem.
Comm. Pure Appl. Math. 44 (1991), no. 7, 819851.
[46] Ni, W.-M.; Takagi, I. Locating the peaks of least-energy solutions to a semilinear Neumann
problem. Duke Math. J. 70 (1993), no. 2, 247281.
[47] Ni, W.-M.; Takagi, I.; Yanagida, E. Stability of least energy patterns of the shadow system for
an activator-inhibitor model. Recent topics in mathematics moving toward science and engi-
neering. Japan J. Indust. Appl. Math. 18 (2001), no. 2, 259272.
[48] Oh, Y.-G. On positive multi-lump bound states of nonlinear Schrdinger equations under mul-
tiple well potential. Comm. Math. Phys. 131 (1990), no. 2, 223253.
[49] Shi, J. Semilinear Neumann boundary value problems on a rectangle. Trans. Amer. Math. Soc.
354 (2002), no. 8, 31173154 (electronic).
[50] Spivak, M. A comprehensive introduction to differential geometry. Vols. 15. 2nd ed. Publish
or Perish, Wilmington, Del., 1979.
[51] Strauss, W. A. Existence of solitary waves in higher dimensions. Comm. Math. Phys. 55 (1977),
no. 2, 149162.
[52] Wei, J. On the boundary spike layer solutions to a singularly perturbed Neumann problem.
J. Differential Equations 134 (1997), no. 1, 104133.
FETHI MAHMOUDI
Facult des Sciences de Tunis
Dpartement de Mathmatiques
Campus Universitaire
2092 TUNIS
TUNISIA
E-mail: fethi.mahmoudi@
fst.rnu.tn
MARCELO MONTENEGRO
Universidade Estadual de Campinas
IMECC
Departamento de Matemtica
Caixa Postal 6065
Campinas - SP
13083-970
BRAZIL
E-mail: msm@ime.unicamp.br
ANDREA MALCHIODI
SISSA
Mathematical Analysis
via Beirut 2-4
34014 TRIESTE
ITALY
E-mail: malchiod@sissa.it
Received May 2008.

Schrödinger Equations

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Schrödinger Equations

Uploaded by

Copyright:

Available Formats

Solutions to the Nonlinear Schrdinger Equation

Carrying Momentum along a Curve

as c tends to 0, and since U

(see (4.50) and the previous formulas) consist-

. This property allows us to solve

; see Proposition 4.14. For tech-

SOLUTIONS TO NLS CARRYING MOMENTUM ALONG A CURVE 1165

; see (4.53) and the previous for-

for a potential frozen at some point in ,

and are related to the kernels of L

s are coordinate vectors for = 1. . . . . n

for all . T = 1. . . . . n, we obtain that

s are coordinate vectors for = 1. . . . . n, we obtain

(s. .) := u(s. . (cs)), u

consisting of approximate (resonant) eigenfunctions of the linearized operator

N, can be identied with the velocity of a curve n(t ) in N, that is equal

) stands for the linear correction in

denote the rst three eigenvalues of (4.28).

is simple, increasing, and differentiable

is 0 for = 0 with multiplicity n, it satises

is simple, radial in ., and radi-

are twice-differentiable in for small, that

to take the sum and integrate; if we choose

) = 1 for all s), then with

the procedure is similar, but notice that in this

(one of them will have multiplicity n 1 by the rotation invariance of the

k), so the smallness of A is equivalent to that of u. Notice that by

and for = 0, we nd that

here stands for the rst eigenvalue of (4.36). We denote by

the components of the relative eigenfunction.

in the second and fourth

in Proposition 4.5), and an afne dependence in c; for 1

in Proposition 4.5). Since the former dependence is more delicate in the

; this is the content of the next result. We

are orthogonal, from the

consist of all the eigenfunctions of

c. For each index

are as in the previous subsection and

decay at the rate e

(see (4.50)), when we replace the

1)1(cs)]) there exist

] with respect to the scalar products induced by the metrics g

V(cs), the function ,

. Moreover, as for (4.93) we can show that

and solve the bifurcation equation as well; this last step

after (4.86). Also, we denote by

) is given by Proposition 4.1, while N

) (see (4.24)) by the

. Using the arguments in the proofs of Proposition 4.9 (see in

, from (6.6) we deduce that

), whose main term is S

(which coincide with A

is odd in z, the products with the even terms will vanish.

. In the latter case,

(in particular , see (5.5)); since ; ranges between

and the last three terms in

, by Fourier cancellation and parity the major terms are indeed A

) solves (4.29) with n

) = 1, and some integration by parts in z, we

are the operators J, T ,

= in the sense that

You might also like